Changelog - AIHubMix

New models

Added claude-haiku-5-5: Anthropic model for high-throughput, cost-sensitive workloads, including summarization, context compression, database queries, classification, and coding sub-agent workflows.

New models

Added gemini-nano-banana-2.1: Google’s efficient image generation and conversational editing model, improving visual quality, prompt adherence, character consistency across turns, and text rendering over Nano Banana 2 while retaining Flash-level speed.

New models

Added doubao-seedream-5-0-flash: ByteDance’s fast image generation and editing model with precise editing, layer separation, and transparent-background output for high-volume image production and rapid interaction.

Image input for Realtime conversations

Realtime conversations with gpt-realtime-2.1 support PNG/JPEG/WebP image input, so you can build conversations combining text and images.
gpt-realtime-2.1 is now available on the production model list for text-and-image Realtime integrations.

2026

September 23

New models

Added claude-opus-5-5: Anthropic’s flagship model for agentic coding, large codebases, computer use, and complex work.
Added gpt-6-sol: OpenAI’s model for complex coding, agent workflows, and long-context reasoning. It supports text and image input, a 1.05M-token context window, and up to 128K output.
Added gpt-6-luna: OpenAI’s efficient model for focused, high-volume tasks. It supports text and image input, a 1.05M-token context window, and up to 128K output.

2026

September 22

New models

Added mimo-v2.6-flash: Xiaomi’s natively multimodal efficient model for frequent calls and large-scale professional tasks.
Added mimo-v2.6-pro: Xiaomi’s natively multimodal flagship reasoning model for complex projects, coding, research, and long-running agent tasks. It supports a 1M-token context window.
Added mimo-v2.6-pro-ultraspeed: Xiaomi’s speed-focused Pro variant, delivering up to 20 times faster output for latency-sensitive use cases.
Added grok-4.7: xAI’s flagship model for coding, agentic tasks, and professional knowledge work, with stronger self-verification and long-context handling.

2026

September 20

New models

Added step-5-preview: StepFun’s next-generation flagship preview for real-world tasks, especially programming and professional knowledge work.

2026

September 18

New models

Added jev-latest: TypeSafe’s structured decision model for fast, typed choices such as routing and classification.
Added qwen3.8-omni-flash: Alibaba Qwen’s native multimodal model for audio/video analysis and agents. It accepts text, image, audio, and video input with a 1M-token context window.
Added glm-5.3-flashx: Z.AI’s high-speed model for coding agents and real-time workflows, with a 1M-token context window and image and video understanding.

2026

September 17

New models

Added union-alpha-free: a free multimodal model for evaluation, research, coding, and agent workflows. It accepts text and image input with a 262K-token context window.

2026

Realtime speech transcription is live (gpt-live-transcribe)

New models

Added gpt-live-transcribe: OpenAI’s realtime low-latency speech transcription model. It outputs text as you speak over a persistent WebSocket connection (delta increments plus completed full text), and is billed by the duration of audio transcribed at $0.017 / minute.

New feature: realtime speech transcription

Added a realtime speech transcription endpoint, based on a persistent WebSocket connection wss://aihubmix.com/v1/realtime?intent=transcription&model=gpt-live-transcribe, supporting both the OpenAI official SDK and native websockets integration.
Audio uses PCM16 / 24kHz / mono, pushed in chunks via input_audio_buffer.append and finalized via input_audio_buffer.commit. It supports configuration for multiple languages (languages), scenario prompt (prompt), keyword biasing (keywords), and noise reduction (noise_reduction).
See the Realtime speech transcription documentation for details.

Gemini streaming tool-call compatibility is improved

For Gemini streaming responses, if a stream has emitted tool calls, the final reason now correctly returns tool_calls even when the last chunk only carries text or the stop marker.
OpenAI-compatible clients that rely on finish_reason to decide whether to execute tools are less likely to treat a tool-call turn as an ordinary completed reply.

Model capability metadata is more complete

/api/v1/models can now expose protocol-specific thinking-content carrier metadata when the model source provides it, helping downstream catalogs and SDKs know which response field contains thinking output.
When the source has not provided this metadata, the field remains absent and existing request behavior is unchanged.

User notification email delivery is steadier

Welcome, quota reminder, invoice, auto-recharge, and channel-block notifications are now separated from verification-code and password-reset email delivery, reducing the chance that business notifications consume authentication email capacity.

Media compatibility updates

Multi-turn reasoning compatibility is improved

Cerebras models now preserve reasoning history more reliably across follow-up turns, reducing 400 errors when clients continue a conversation or complete tool loops.
Reasoning-history handling is more consistent across Chat, Responses, and Messages, without changing behavior for other providers.

MAI image editing is steadier

Legacy /v1/images/edits requests routed to MAI image models now use the edit endpoint instead of the generation endpoint, preventing image edits from being forwarded to the wrong provider path.
MAI generation and edit path configuration is clearer, making mai-image-2.5, mai-image-2.6, and Flash edit flows easier to validate and troubleshoot.

Media failure messages are more actionable

Image, video, and async media tasks now show clearer user-facing messages for provider unavailable, model-not-enabled, and deployment-not-found cases, while preserving existing error codes, HTTP statuses, retry behavior, and billing behavior.
Immediate responses, task detail/list responses, and webhooks now present failed media tasks more consistently so support can keep using the request or task ID for diagnosis.

Responses image-tool billing and Azure deployments are steadier

Image-tool calls are more stable

When using the image_generation tool through Responses, Azure image deployments now respect the deployment name configured in model mapping, including dotted names such as gpt-image-2.5-*, reducing 404s caused by rewritten deployment IDs.
If multiple image tools map to the same deployment but imply different billing models, the request now receives a clear error instead of producing ambiguous consumption records.

Image consumption records are easier to reconcile

Token-priced image-tool usage now keeps the text-input, image-input, and image-output token basis in consumption records, making user-side usage easier to compare with provider bills.
Free image models explicitly configured with a zero rate continue to settle as free instead of falling back to per-image billing.

Qwen 3.8 27B: New models

New models

Added qwen-3.8-27b: Alibaba’s open-source dense vision-language model with native understanding of text, images, and video. Built on the Qwen3.5 architecture, it improves coding, professional and research tasks, and long-running complex workflows. It supports a 131K-token context window and flexible cognitive control.

GPT Image 2.5: Models and media improvements

New models

Added gpt-image-2.5-flare: OpenAI’s image model for fast, everyday high-quality image generation. It accepts text and image input, produces image output, and supports quality settings low, medium, high, xhigh, max, and auto.
Added gpt-image-2.5-sunburst: OpenAI’s image generation and editing model for workflows requiring high editing precision. It accepts text and image input, produces image output, and supports quality settings low, medium, high, xhigh, max, and auto.

Image billing fallback is more reliable

When upstream image usage omits InputTokensDetails, available input usage is still used for billing instead of being recorded as zero; consumption records and related logs use the same fallback.
Complete upstream usage details remain preferred when available.

Public model information is more consistent

Public model-information responses now expose only supported public fields, and retired integrations no longer appear in public data.
Performance metrics are written back from measured results, reducing stale values in model details and API responses.

Media errors are clearer and more consistent

Image, video, and async media endpoints now use consistent public error codes, HTTP statuses, and actionable messages for invalid input, unsupported formats, provider rejection, unavailable models, blocked output, and expired results.
List, detail, and webhook responses use the same error semantics without exposing raw provider responses or media payloads.

Playground endpoint filtering is more accurate

Model endpoints are now filtered by protocol version, preventing entries for other versions from appearing in a selected Playground/API list.
This reduces protocol-mismatch failures when selecting a model or endpoint.

Legacy Claude system-role requests are more compatible

Older Claude models that still use a system role are now handled more consistently across Chat and Responses requests.
The compatibility path preserves existing message, cache, and thinking behavior, so callers do not need a separate rewrite for legacy models.

DeepSeek V4.1 Flash: Models and platform improvements

New models

Added deepseek-v4.1-flash: DeepSeek’s internal-beta model built on a new architecture with native multimodal support. It accepts text and image input, produces text output, and supports reasoning, tool calling, and structured outputs. Trial access is rate-limited and intended for testing only, and is not recommended for production.

Gemini billing is safer when usage is missing

For native Gemini passthrough calls, when the client disconnects or the provider does not return usage, AIHubMix now estimates delivered output usage from the content already sent back instead of treating the pre-bill value as final billing.
Web Search and cachedContent keep their existing billing semantics; when provider usage is available, it still takes priority.

Model routing errors are easier to diagnose

When a model does not exist or no compatible integration is available, the error message now echoes the original model ID from the request and points users to the model page for the correct ID. Mapping, redirects, and fallback no longer hide the model name the user actually sent.

GPT and Claude tool-call compatibility is steadier

Reasoning-model Chat requests bridged to Responses now filter sampling parameters such as temperature and top_p when the target model does not accept them, reducing 400 errors for models such as gpt-5.6 and gpt-6-astra.
AWS Bedrock Claude requests remove the advisor tool declaration that Bedrock does not currently support, while preserving other tools, non-AWS behavior, and message content.

Temporary file upload API added

Added POST /v1/oss/upload for API-key authenticated uploads of one image, PDF, text, audio, or video file up to 50 MB, returning a read URL valid for 15 minutes.
The upload endpoint is not billed separately and is intended for temporarily hosting multimodal files before passing them to models. File-type validation is not a virus scan or a guarantee that every model accepts the file.

GPT: Tools and prompt caching

Model badges are more consistent

The model catalog, model detail pages, and ?tag=new lists now derive the New badge from model launch time. New expires automatically after the configured window, and the old Hot badge is no longer shown. Recently launched parent models also keep their New badge when supplier self-binding data exists.

GPT tool-call compatibility is improved

gpt≥5.5 models called through /v1/chat/completions with tools now use the Responses compatibility path automatically, reducing upstream 400 errors. Claude Messages bridge calls for these GPT tool workflows follow the same compatibility path.
OpenAI and Azure direct channels filter incompatible reasoning parameters before sending requests, avoiding failures caused by provider-specific parameter dialects.

GPT cache-write billing tiers are more accurate

GPT requests that include cache writes now choose billing tiers from the original total input length, which better matches official long-context pricing behavior. Normal input counts, cache-write quantities, and non-GPT billing behavior are unchanged.

GPT-6 Astra / MAI Image 2.6: New models

New models

Added gpt-6-astra: OpenAI’s model for complex reasoning, coding, computer use, research, and document creation. It supports a 1.05M-token context window, up to 128K output, low-to-max reasoning effort, asynchronous tool calling, and mid-turn steering. Launch announcement.
Added mai-image-2.6: Microsoft’s image generation and editing model, supporting text-to-image generation and image-guided editing in multiple aspect ratios.
Added mai-image-2.6-flash: Microsoft’s low-latency version of MAI-Image-2.6, supporting image generation in multiple aspect ratios and image-guided editing.

Image generation reliability

Image generation requests are more stable

Image generation, edits, variations, and gpt-image-2 prediction requests now use Azure rate-limit headers together with distributed in-flight limits. Under high concurrency or upstream 429 responses, AIHubMix can throttle earlier, release capacity, and try available channels more consistently.
For by-bill image models, missing upstream usage now falls back to a consistent usage value, keeping rate limits and billing aligned. If the Redis limit component is unavailable, image requests fail conservatively instead of sending unbounded traffic upstream.

Media task errors are more consistent

Media task failures across /ai/v1, /call/media, task detail, task lists, and webhooks now render from the same public error codes and failure snapshots. Clients should see more consistent HTTP statuses, error.code values, and failure reasons for retries, UI display, and support workflows.
Upstream raw text, media URLs, data URIs, Base64 payloads, and internal error details stay out of user-visible responses. Bad JSON, unknown models, oversized input, and upstream rejection are mapped to clearer public errors.

Long-running agents and multimodal models

New models

Added muse-spark-1.3: Meta’s multimodal reasoning model for long-running agentic, multi-agent, and complex coding workflows. It supports text, image, audio, and video input, text output, a 1M-token context window and up to 1M output, reasoning, and tools. Compared with Muse Spark 1.2, it completes long-horizon agentic and coding work more efficiently with fewer unnecessary turns and tool calls and more concise output.
Added gemini-3.8-flash: Google’s most intelligent Flash-series model for long-running software engineering, autonomous agents, and complex enterprise workflows. It supports text, image, audio, and video input, a 1M-token context window, up to 65K output, reasoning, web access, tool calling, structured outputs, and long context.
Added gemini-3.8-flash-free: a free trial version of Gemini 3.8 Flash with the same core multimodal and agent capabilities. Resources are limited, stability is not guaranteed, and 429 errors may occur; use gemini-3.8-flash for production workloads requiring stable concurrency.

Claude Fable / Qwen / MiniMax: New models

New models

Added claude-fable-5-1: Anthropic’s flagship model for demanding reasoning, long-running agents, multi-step research, and document, spreadsheet, and slide work. It supports text and image input, a 1M-token context window, thinking, tool calling, code interpreter, computer use, memory tool, structured outputs, citations, and prompt caching, with cached-input reads priced at one quarter of Claude Fable 5.
Added mercury-2.5-preview: Inception’s preview diffusion LLM for high-speed text generation. It supports a 260K-token context window, and the published model page highlights parallel token generation and optimization with up to 1,107 tokens/s on standard GPUs. Capability flags are not fully published yet, so validate target capabilities with a minimal call before relying on them.
Added qwen3.8-max-2026-09-02: a September 2 snapshot of Qwen3.8 Max for complex engineering projects, long-horizon autonomous development, and multi-tool agent collaboration. AIHubMix currently lists text input and roughly 991K tokens of context; capability flags are not fully published yet, so verify vision, tool, or structured-output needs before production use.
Added minimax-h3-max: a MiniMax H3-family multimodal creation model for advertising, branding, e-commerce, product design, UI/UX, and game workflows. AIHubMix currently lists text and image input; H3-family material highlights video, audio, and reference-editing workflows, while capability flags are not fully published yet, so validate the target path before production use.

Model availability and media input

Retired-model errors are clearer

Calls to retired models now return an explicit model_retired 404 instead of a generic model-not-found response. When a successor model is configured, response headers include retirement and successor information to help clients migrate.
Key creation or editing rejects retired models, and smart routing avoids retired models, reducing failures caused by requests landing on unavailable models.

Media input errors are more actionable

Unsupported image, video, and other media inputs now consistently return media_form_unsupported, with the detected media type and the input forms accepted by the selected model.
Media error messages are aligned across sync, async, and webhook flows, while URLs, Base64 payloads, and internal causes stay hidden from user-facing responses.

Google image input validation is temporarily higher

AIHubMix gateway validation for Google/Gemini image inputs is temporarily raised to 20 MB, so larger image requests are no longer blocked by the platform before upstream submission.
The current path still uses Google’s inline media mode, so Google may still reject inputs above its effective accepted range. Watch upstream errors when sending large files.

Hy4 Preview: New models

New models

Added hy4-preview: Tencent Hunyuan’s 770B-parameter MoE model with 49B active parameters, optimized for Agent, coding, and productivity workflows. It strengthens task understanding, planning, tool use, and sustained execution across complex multi-step tasks, software engineering, document processing, information analysis, office automation, web generation, and cross-tool collaboration. It supports a 1M-token context window, up to 64K output, reasoning, web search, tool/function calling, and structured outputs.

Qwen 3.8 Flash: Models and MCP image tools

New models

Added qwen3.8-flash: Alibaba Cloud Qwen’s flagship native vision-language model for coding, office tasks, long-context reasoning, and agent workflows. It supports a 1M-token context window, up to 128K output, thinking, web access, tool calling, and structured outputs. Compared with Qwen3.7-Plus, it strengthens coding and office capabilities while improving training and inference efficiency.

MCP image-tool calls are clearer

The MCP image generation tool now requires callers to pass a model explicitly, avoiding silent fallback to a default model when none is provided.
openai/gpt-image-1 is the standard model ID shown and forwarded externally. The historical typo remains accepted as an input-only compatibility alias and is normalized before the request is sent.

Broadcast, remote MCP and Access Key

New models

Added glm-5.3-flash: Z.AI’s high-efficiency multimodal model for coding agents, complex reasoning, and long-horizon software engineering tasks. It supports text, image, and video inputs, roughly 1 million tokens of context, up to 128K output, thinking, tool calling, and structured outputs. Built on the GLM technology stack with further post-training and optimization, it emphasizes inference efficiency and responsiveness.

New feature

Added Broadcast, which automatically sends traces from AIHubMix API requests to LangWatch without requiring additional application instrumentation. It supports privacy mode, sampling rates, and API key filtering, with model, token usage, actual billed cost, latency, session, and trace data available in LangWatch.

MCP and Access Key entry points are more complete

The legacy /mcp/ entry point can keep using the existing authentication flow while routing to the unified remote HTTP MCP capabilities. Older clients do not need to migrate immediately and can still access model and pricing lookup, document search, credit checks, chat, image, and video tools.
The console will provide a read-only Access Key copy path, and external MCP clients can use the X-Aihubmix-Access-Key header. The server does not rotate the key or cache this response.

Transparent-background image requests are steadier

gpt-image-2 requests with transparent background now prefer OpenAI-native channels that support the capability, reducing repeated failures on channels that do not. Other image request paths are unchanged.

Agent-facing files update more quickly

Model lists and guides for AI Agents and external tools now refresh more promptly.

API migration guidance

Legacy media API migration errors are clearer

For media models that no longer support legacy protocol access, /v1/images/generations, /v1/images/edits, /v1/models/:organization/:model/predictions, and /v1/videos now return protocol_not_supported with guidance to the /ai/v1 async media APIs and related documentation.
model=auto keeps the existing automatic model-selection behavior. If the final selected model does not support the legacy endpoint, callers receive a clear error instead of being silently switched to another model. Playground also hides legacy entries for these models to reduce accidental use.

Wan 3.0 / GPT-5.6 Sol: Models and pricing updates

New models

Added wan3.0-video: Alibaba Cloud Wan’s preview All-in-One video model unifies text-to-video, first/last-frame image-to-video, and reference-video workflows, supporting up to 30 seconds at 30 fps in 480P/720P/1080P for integrated generation and editing.
Added wan3.0-video-prime: a capability-aligned high-speed edition for integrated video generation and editing workflows, with faster end-to-end turnaround.

GPT-5.6 Sol price reduction

gpt-5.6-sol has synchronized the official 20% price reduction and is now priced at $4 per million input tokens and$ 20 per million output tokens. Existing request paths are unchanged, while complex reasoning, coding, and long-context workloads cost less to run.

Async media task enablement is clearer

When an account has not enabled async tasks, media creation APIs now include the matching console enablement page in the async_not_enabled error. aihubmix, inferera, and shkq users are guided to their own console, reducing dead-end 403 responses.

DeepSeek V4: Models and media improvements

New models

Added deepseek-v4-flash-0731-fast: a high-speed variant of DeepSeek’s agentic model for coding, tool use, and high-volume agent workloads. It preserves V4 Flash 0731 capabilities while delivering faster inference; compared with V4 Pro, it prioritizes lower latency and execution efficiency. It supports a 1M-token context window, up to 384K output, thinking, tool calling, and structured outputs.
Added deepseek-v4-flash-vision-exp: DeepSeek’s experimental multimodal visual-understanding model with text and image input and text output. It is suited to image description, screenshot OCR, chart analysis, and visually grounded agents; its pure-text agent, reasoning, and world-knowledge capabilities stay in line with the official DeepSeek V4 Flash release.
Added ox-alpha, a stealth reasoning model from an anonymous third-party provider whose true developer and owner have not been disclosed. Free on AIHubMix, it supports text and image input, text output, and a 1,048,576-token context window for coding, long-horizon software engineering, sustained agentic work, complex reasoning, and text-plus-visual workflows.

Image generation failure recovery is steadier

When an image generation request hits a temporary service error, AIHubMix now recovers more consistently. If the error clearly reflects a capability limit of the selected model, it still stops retrying and returns the reason. This should reduce direct failures caused by transient service differences for models such as gpt-image-2.

Agent integration and model discovery

Platform announcement

Launched agent-facing access capabilities: the agents.md onboarding guide, site-wide and per-model llms.txt machine-readable documentation, the Playground Skill (distributed at a fixed URL), and Playground deep links (?models= opens up to 6 model tabs from one link, ?config= loads a configuration directly). To view two models’ specs and pricing side by side, use the compare page at https://aihubmix.com/compare/{model_a}/{model_b}. See Agent Access for the full list of entry points.

Image generation API parameters and errors are more consistent

Image generation requests now handle size parameters such as resolution and imageSize more consistently. When a request is outside the current model capability, the API returns clearer errors so callers can fix parameters faster.

Model Hub now supports main models with versions

Related variants such as GLM 5.2, the Coding edition, and the Free edition can be presented as one main model with multiple versions. This makes browsing, price comparison, and version selection easier to understand.

Async image tasks can move faster

Async image tasks can now progress with bounded concurrency based on available capacity. During busy periods, queues should clear more smoothly and task status/result delivery should be steadier.

GPT-5.6 Sol: Models and media improvements

New models

Added gpt-5.6-sol-disc: AIHubMix’s limited-time discount route for OpenAI’s GPT 5.6 Sol, offering up to 50% off with the same underlying capabilities as standard gpt-5.6-sol. This frontier reasoning model is suited to complex coding, professional knowledge work, deep research, and long-running agents, with text and image input, about 1.05M tokens of context, up to 128K output, thinking, tools, and structured outputs.

Media generation tasks are more reliable

Completion time, result expiry, and artifact archival are now more consistent for image and video tasks. During rolling releases or high concurrency, task status, downloadable results, and user-facing records are less likely to be delayed or processed twice.
Media production capacity has been increased, and larger media responses have more headroom, improving availability for long-running or large-artifact tasks.

Model information and media compatibility

Model details are more complete

Model details now expose release_date, so the official release date is easier to distinguish from AIHubMix launch time across model pages and comparisons.

Gemini image and upload compatibility is steadier

Gemini image generation/editing and Vertex multi-image flows now preserve imageSize and related image parameters more consistently, reducing size or protocol mismatches.
Gemini file uploads keep the original Content-Type, so videos and other media are less likely to be misclassified.

Public errors are easier to read

User-facing schema, enum, and capability errors now surface clearer messages, making parameter issues easier to diagnose.

GLM-5.3: New models

New models

Added glm-5.3, Z.AI’s production API for its text-only flagship reasoning model, built for complex software engineering, long-horizon Agent tasks, and vulnerability analysis. It supports a 1M-token context window, up to 128K output, and unlimited concurrency on AIHubMix, with a limited-time 10% discount.

MAI Thinking 1: New models

New model

Added mai-thinking-1: Microsoft MAI-series reasoning model for enterprise-scale reasoning, math, general intelligence, and high-throughput workloads. It supports a 256K-token context window, Chat Completions, and structured outputs, making it suitable for long-context reasoning, stable response formats, and cost-conscious always-on usage.

Gemini 3.7 / Coding GLM-5.3: API compatibility and task reliability

New models

Added gemini-3.7-flash: Google’s natively multimodal reasoning model for coding, agents, web development, and knowledge work. It supports text, image, audio, and video input, a 1M-token context window, adjustable thinking levels, tool calling, web search, and structured outputs, with stronger coding, tool use, multi-step planning, and instruction following than Gemini 3.6 Flash.
Added coding-glm-5.3: Z.AI’s reasoning model for coding and agentic workflows, designed for complex software engineering, long-running agents, and vulnerability analysis. Built on the same base model as GLM-5.2, scaled post-training improves coding, task execution, and token efficiency. This limited-time preview is for testing and evaluation only; service stability is not guaranteed and production use is not recommended.

API compatibility and task reliability

Native Gemini embedding requests now preserve the model-service source for the fileUri returned by the Gemini Files API. When embedContent or batchEmbedContents references uploaded files, 403 errors caused by source mismatches are reduced; batch requests that mix files from different sources now return a clear 400.
When OpenRouter returns a structured model provider error, the API error now prioritizes the inner message, param, and type, reducing long, hard-to-parse messages. Single-provider errors are also less likely to affect the whole OpenRouter service.
Media async task polling, terminal archival, and artifact handling are more reliable. Long-running image and video tasks are less likely to be processed twice, lose terminal state, or be delivered late during high concurrency, restarts, or large artifacts; public API fields, status words, and billing formulas are unchanged.

Grok 4.6 / DeepSeek V4 Pro: New models

New models

Added grok-4.6: xAI’s flagship multimodal reasoning model for coding, long-running agents, knowledge work, and interactive application development. It supports image understanding, a 500K context window, tool calling, and structured outputs, with stronger multi-step execution, self-verification, coding, and visual project generation than Grok 4.5.
Added deepseek-v4-pro-0813: DeepSeek’s high-performance general-purpose reasoning and agent model for complex reasoning, coding, long-document analysis, and agentic workflows. It supports thinking and non-thinking modes, a 1M-token context window, up to 384K output, tool calling, and the Responses API.

Agnes: Models and media improvements

New models

Added agnes-2.5-flash, agnes-2.5-pro, agnes-2.5-pro-alpha, and agnes-image-2.1-flash. Text and image Agnes models are now available through the unified AIHubMix entrypoint.

Improved Azure Responses tool-call compatibility

When using namespace tools through Azure Responses, blank or missing descriptions no longer cause Azure to reject the request with a 400 before model execution. Tool calls that already worked through the official OpenAI interface should behave more consistently on Azure.

More complete model and supplier page data

/call/devs now supports a supplier website field, so model aggregation and supplier pages can show an Official site entry once data is filled in. /providers/<slug> pages also now preserve edge-cache behavior for steadier page responses.

Asynchronous tasks and protocol compatibility

Gemini interactions now supports async tasks

Gemini interactions requests can now explicitly use background: true for async tasks, then use the returned task ID to query status, cancel, or clean up the task. Long-running multimodal/omni interactions are easier to integrate without keeping a synchronous connection open until completion.
Async tasks are settled against real usage, with token columns, full release on cancellation, and large-response safeguards filled in. Console usage records, balance pre-deduction, and final charges should stay more consistent.

Improved Claude Messages thinking compatibility

When the Messages protocol calls OpenAI-compatible model providers, returned thinking/reasoning content is preserved and can be sent back in later turns. Apps using Claude SDKs, multi-turn tools, or DeepSeek thinking mode should see fewer empty responses, 400 errors, or lost reasoning context.
Unsigned thinking blocks are handled more compatibly when traffic switches to native Claude service, reducing format errors in multi-model fallback or bridge flows.

Improved Doubao video prompt and reference compatibility

When calling Doubao video models, prompt and reference assets in extra_body.content can now work together; requests that contain only assets keep both the prompt and the asset order.
Create responses now echo size and prompt information closer to what is actually sent, while still ignoring unsupported native fields. This reduces video-generation failures caused by extra SDK parameters.

Usage billing and log accuracy

More complete cumulative usage for media generation

After a media-generation task is settled, account-level cumulative used quota is updated as well. In the console, cumulative usage, balance deduction, and usage records reflect real media-generation consumption more completely.

Improved Azure model-name compatibility

When calling Azure models whose deployment/model name contains dots, the model name is preserved across the Chat, Responses, and Messages conversion paths and is no longer rewritten automatically. Existing Azure call behavior remains unchanged.

Seedance 2.5: Video generation updates

New video generation model

Added doubao-seedance-2-5-260628: ByteDance Seed’s next-generation unified multimodal audio-video generation model. It creates up to 30 seconds of synchronized audio and video in one pass, supports multi-round extensions, and improves storytelling, transitions, reference support, realism, and precise editing for filmmaking, advertising, education, simulation, and long-form content.

Wan 3.0 / Qwen Image 3.0: Models and media improvements

New models

Added wan3.0-video: Alibaba Tongyi Lab’s all-in-one video generation and editing model, currently in public beta. It supports native videos up to 30 seconds, production-grade character consistency, lifelike visuals and sound, and unified reference, editing, replication, and motion-driving workflows for advertising, e-commerce, filmmaking, animation, video editing, and document-to-video creation.
Added qwen-image-3.0: image generation and editing model from Alibaba Cloud’s Qwen team, supporting text-to-image, reference-based creation, and image editing. It balances image quality and speed for social media, e-commerce, creative design, everyday content production, and frequent large-scale creation.
Added qwen-image-3.0-pro: Alibaba Cloud Qwen’s flagship image generation and editing model for advertising, brand visuals, UI, presentations, product imagery, and professional design. Strong at complex layouts, accurate Chinese and English text rendering, realistic materials, reference-based editing, detail, composition, and commercial-grade visual quality.

Gemini / Seedance: Multimodal embeddings and video compatibility

Gemini embedding supports multimodal input and per-modality billing

gemini-embedding-2-preview can now handle text, image, audio, video, and document inputs through both the OpenAI-compatible embeddings endpoint and the native Gemini path. Users building multimodal retrieval, content understanding, or knowledge-base workflows can use more input types on the same model.
Gemini embedding charges now prioritize the per-modality token usage returned by the model provider. Image, audio, video, and document inputs are no longer treated as a single text estimate, so multimodal embedding bills should better match actual usage.
Single-input embedding requests can call the newer Gemini embedding path directly. Multi-input batch requests now return a clear capability-mismatch response.

More reliable Seedance 2.0 video-generation calls

Seedance 2.0 references for images, audio, and video are mapped more accurately by media type, with clearer constraints for adaptive, 4K, and size options. Video-generation workflows should see fewer failures caused by parameter mapping mismatches.
Adaptive-duration requests are pre-deducted at 15 seconds and settled against the actual duration after the task finishes, keeping balance pre-deduction and final billing more consistent.
When media generation fails, APIs now return sanitized error details, making it easier for users and support teams to diagnose parameter, signed-URL, or model-provider limit issues without exposing sensitive credentials.

Qwen 3.8 Max: Models and provider availability

New models

Added qwen3.8-max: Alibaba Cloud’s flagship native vision-language model, built on a 2.4-trillion-parameter Mixture-of-Experts (MoE) architecture and supporting context windows of up to 1 million tokens. It is well suited for complex multimodal understanding, advanced reasoning, software development, agentic workflows, and long-context processing. At a similar price to Qwen3.7-Max, Qwen3.8-Max delivers significant improvements in reasoning, coding, and agent capabilities, with overall performance comparable to today’s leading models.

Clearer supplier availability in model details

Model details now show suppliers that are better suited for current requests first, while backup-only or temporarily unavailable suppliers appear later. When choosing a model supplier or troubleshooting availability, users and support teams can more quickly see which suppliers are currently the better choice and avoid decisions based on stale ordering.

Usage billing and log accuracy

More accurate small-request billing and usage logs

Small requests on by-bill models no longer disappear from billing and usage logs because the calculated amount rounded down to zero. Affected requests are now charged at the minimum effective quota, keeping invoices, balance pre-deduction, and usage records aligned; support and operations teams should see more complete records for high-frequency embedding-style traffic.

DeepSeek / MiniMax / GPT: Models and pricing updates

New models

Added deepseek-v4-flash: an efficiency-optimized Mixture-of-Experts model with 284B total parameters, 13B active parameters, and a 1M-token context window. It is designed for fast inference and high-throughput workloads while retaining strong reasoning and coding performance, making it suitable for coding assistants, chat systems, and agentic workflows.

New video generation model

Added minimax-h3: MiniMax’s general-purpose multimodal video model. It accepts text, image, audio, and video inputs, and supports text-to-video, image-to-video, first/last-frame generation, reference-based generation, and video editing. Users building video creation, ad creative, product demo, or multimodal content workflows can call it through the AIHubMix unified API.

GLM-5.2 limited-time time-of-day offer

glm-5.2 is discounted by daily time window, in UTC: 50% off from 14:00 to 24:00 and 30% off from 00:00 to 14:00 each day. Limited-time offer.

GPT-5.6 Luna/Terra price reductions

gpt-5.6-luna has synced the official 80% price cut and is now priced at $0.20 input and$ 1.20 output; gpt-5.6-terra has synced the official 20% price cut and is now priced at $2.00 input and$ 12.00 output. Users with steady traffic on these models can re-evaluate model choices for cost-sensitive and general workloads.

AIHubMix: Official MCP service

AIHubMix official MCP Server is available

Added the official hosted MCP endpoint mcp.aihubmix.com/mcp or mcp.inferera.com/mcp for MCP-capable clients. Developers can use one entrypoint from tools such as Claude Code, Codex, and Cursor to query models and pricing, search docs, check credits, and call chat, image-generation, and video-generation tools.
MCP credentials are passed by the client on each request and are not stored by the server, making the endpoint suitable for developers who want direct AIHubMix model and media access inside IDEs or agent tools.

Jina Reranker V3.5: New models

New model

Added jina-reranker-v3.5: Jina AI’s 0.6B multilingual listwise document reranker and a same-schema upgrade from jina-reranker-v3. It targets RAG, search, and structured-data ranking workflows, supports long-context candidate ranking, and improves domain robustness, multilingual retrieval, structured ranking, and inference efficiency.

Caching, media tasks and data refresh

More accurate cache billing

For explicit and implicit cache requests, AIHubMix now identifies cache read/write usage more consistently and prioritizes the cache type returned by the model provider. Users of cache-capable models such as Qwen and Claude should see billing details that better match actual usage.

More reliable media generation and task results

The first media-generation model list has been narrowed to verified models, and an issue that could return 404 for some image-model calls through the media endpoint has been fixed. Console downloads for LLM tasks now also cover text archives, making large historical results easier to retrieve and deliver.

Fresher website and model data

Cache behavior for the website, model data, blog data, and sitemaps has been adjusted so releases take effect more reliably. Model pages and documentation entrypoints are less likely to be affected by stale cached data after updates.

Unified asynchronous media API

Media generation APIs are available

Added /ai/v1/images/generations and /ai/v1/images/edits media-generation entrypoints with unified async tasks, result queries, artifact downloads, and webhook callbacks. Image and video workflows can now integrate through the same task lifecycle, making long-running generation results easier to track.
Generated artifacts are returned through AIHubMix task results and support limited downloads and bulk retrieval. Customer-facing teams should watch early feedback on task status, download-link validity, and billing details.

Claude Opus 5: New models

New models

Added claude-opus-5: Anthropic’s flagship model for demanding reasoning, coding, and long-horizon agentic work. It excels at end-to-end software tasks, code review and bug finding, chart and document analysis, complex office deliverables, and parallel subagent coordination, with a 1M context window and up to 128K output.

Qwen TTS / MAI Image: Audio generation and model updates

New models

Added mai-image-2.5-pro: Microsoft’s flagship image generation and editing model, featuring high-fidelity realism, precise in-image text rendering, and powerful editing for commercial design, product photography, and professional creative workflows.
Added qwen-audio-3.0-tts-flash: Qwen’s real-time speech synthesis model with expanded language and Chinese dialect support, expressive voice control, and first-packet latency under 200 ms for voice assistants, real-time dialogue, and intelligent customer service.
Added qwen-audio-3.0-tts-plus: Qwen’s premium speech synthesis model with richer control of emotion, tone, character, speaking rate, volume, and style, designed for content creation, audiobooks, dubbing, brand voices, and other high-quality speech workflows.

Qwen TTS audio-generation capabilities

Both Qwen TTS models can be called through /v1/audio/speech; non-streaming requests return an audio result URL, while streaming requests return SSE audio-generation events. The table below combines AIHubMix-compatible fields with Alibaba Cloud’s official Qwen-Audio-TTS docs; voice must be selected per model, and plus/flash system voices are not interchangeable.

Capability	`qwen-audio-3.0-tts-plus`	`qwen-audio-3.0-tts-flash`
Input text `input`	Required, up to 4096 characters through the AIHubMix-compatible endpoint. Text can include emotion tags such as `[excited]` and `[laughing]`.	Required, up to 4096 characters through the AIHubMix-compatible endpoint. Text can include emotion tags such as `[excited]` and `[laughing]`.
Voice `voice`	System voices: `longanlingxin`, `longanlufeng`; also 500+ base/cloned voices.	System voices: `longanhuan_v3.6`, `longjielidou_v3.6`, `loongeva_v3.6`, `loongjohn`; also 500+ base/cloned voices.
Audio format `response_format`	Default `mp3`; also supports `wav`, `pcm`, and `opus`. `aac`/`flac` are not supported.	Default `mp3`; also supports `wav`, `pcm`, and `opus`. `aac`/`flac` are not supported.
Speed `speed`	Supports `0.5` to `2.0`, default `1.0`.	Supports `0.5` to `2.0`, default `1.0`.
Style instructions `instructions`	Natural-language control for role, emotion, and accent, such as “use a gentle tone”.	Natural-language control for role, emotion, and accent, such as “use a gentle tone”.
Streaming `stream_format`	Omit for non-streaming URL JSON; set `sse` for streaming events.	Omit for non-streaming URL JSON; set `sse` for streaming events.

Improved Messages response compatibility

/v1/messages responses from bridged models and Claude-compatible services such as Bedrock now align more closely with the Anthropic Messages ecosystem, reducing errors in strict SDKs or schema validators caused by missing response fields.

More complete usage after AWS Bedrock stream disconnects

When a client disconnects from an AWS Bedrock streaming request, AIHubMix continues collecting the model inference provider terminator and final usage, reducing mismatches between request logs, billing details, and model inference provider usage in disconnect scenarios.

Qwen3-Coder / Gemini: New models

New models

Added qwen3-coder-480b-a35b-instruct: Qwen3-Coder’s flagship coding model for code generation, software-engineering agents, and tool-calling workflows, suited for long-context code understanding and automated development tasks.
Added gemini-2.5-flash-lite: a lightweight Gemini 2.5 model for low-latency, lower-cost high-frequency text, multimodal, and batch workloads.
Added Qwen/Qwen3-235B-A22B-Instruct-2507: the Qwen3 235B-A22B Instruct 2507 release for general chat, instruction following, multilingual work, and long-context tasks.

Gemini / GLM / Qwen: Models and platform improvements

New models

Added gemini-3.6-flash: Google’s latest Flash model for complex workflows, computer use, chart reasoning, and long-context tasks.
Added gemini-3.5-flash-lite: a lighter, lower-cost Gemini 3.5 model for high-throughput, everyday automation, and batch processing.
Added glm-5.2-fast-preview: Zhipu GLM-5.2 fast preview model for low-latency chat, fast reasoning, and high-throughput applications.

More reliable Qwen 3.8 streaming

Streaming calls to qwen3.8-max-preview through /v1/responses and /v1/messages are more reliable. Valid streams that could previously surface as empty_stream or 502 can now complete normally.

Gemini / Responses: Cross-protocol compatibility

Gemini image parameters passthrough

When calling Gemini or Vertex image-output models through the OpenAI-compatible Chat Completions API, clients can now pass aspectRatio and imageSize in extra_body.generationConfig.imageConfig.

More accurate Gemini embedding billing

When Gemini embedding responses include real usageMetadata, AIHubMix now bills from the token usage returned by the model provider first, while keeping the previous local estimate as a fallback when usage is missing.

Improved Responses dynamic tool compatibility

/v1/responses requests can now carry message-local function tools on system/developer messages, and those tool declarations are preserved when bridged to Chat models. Clients using dynamic tool workflows with models such as Kimi K3 should see more reliable tool calls.

Responses stop parameter fixed

For /v1/responses calls that are bridged to Chat models, stop is now forwarded to the model provider. A single stop sequence can take effect normally, while stop counts or lengths beyond model provider limits return boundary errors consistent with Chat Completions.

DeepSeek V4 time-based billing support

DeepSeek V4 models can now support configured peak/off-peak billing. Pre-deduction and final billing use the same request submission time, so streaming requests that cross a time boundary are still charged by submission time. Pre-deduction protection is limited to models configured for time-based billing, avoiding balance-freeze changes for ordinary models.

Claude: Streaming reliability and compatibility

Claude Messages thinking compatibility

When /v1/messages requests are bridged to OpenAI or Azure Chat models, Claude thinking settings are now mapped to a reasoning effort supported by the model provider, reducing 400 errors for thinking-enabled requests. Native Claude thinking requests such as Opus 4.8 remain compatible.

More stable long streaming requests

The production gateway stream limit has been raised from 1 hour to 2 hours. Long-thinking and long-generation workloads such as Kimi K3 are less likely to be cut off while reasoning or generating.

Qwen 3.8 Max Preview: Models and pricing updates

Added qwen3.8-max-preview:
Limited-time offer: Qwen3.8-Max-Preview is billed at just 10% of the standard price, effectively 10× your usage. Limited time, first come, first served.

Qwen3.8-Max-Preview is the latest-generation foundation model in the Qwen family, with 2.4T parameters and still evolving. Compared with the previous flagship Qwen3.7-Max, it delivers major gains in core capabilities such as Coding and Cowork (professional productivity), and shows world-leading overall performance on complex, long-horizon tasks like full-stack development, data analysis, and Office workflows.

Kimi K3 / Hunyuan 3D: Models and platform improvements

New models

Added kimi-k3: Kimi K3 is Moonshot AI’s open 3T-class long-context model, with 2.8T parameters, a 1M-token context window, and native vision input support. It is suited for long-horizon coding, knowledge work, complex reasoning, and multimodal understanding. See the Kimi K3 hands-on guide for usage (three-API examples and a capability support matrix).
Added hy-3d-3.1: Tencent Hunyuan 3D Professional model for text-to-3D, image-to-3D, and multi-view 3D generation. It is suited for game assets, ecommerce presentation, 3D printing, and product-design workflows; the 3.1 model improves geometry and texture quality and supports eight-view input.

3D generation API is available

Added the asynchronous /v1/3d/generations API, initially connected to Tencent TokenHub Hunyuan 3D. It supports task submission and result querying, making 3D asset generation easier to integrate into automated workflows.

Kimi K3 dynamic tool compatibility

Message-level dynamically loaded tools in Kimi K3 requests are now preserved and forwarded, so model providers no longer return 400 errors caused by dropped tool declarations. Clients that use Kimi dynamic tool loading should see better compatibility.

Clearer request-conversion errors

Compatibility has improved for Chat Completions custom tools, Cohere text content blocks, and requests with assistant.content: null. Valid requests reach model providers more reliably, while malformed or unsupported inputs now return structured errors instead of being misreported as empty HTTP 200 successes.

Updated Seedream image billing

doubao-seedream-5-0-pro image generation now supports billing by output pixel tier and input-image count, so text-to-image and image-to-image requests are charged closer to their actual generation specs.

Gemini: Tool calling improvements

Improved Gemini tool-call compatibility

When calling Gemini or Vertex through the OpenAI-compatible API, requests whose tool definitions or structured-output schemas contain empty-string enum values no longer fail model provider validation with a 400. Tool-calling and JSON Schema workflows are now more reliable.

OpenAI: Audio generation and model updates

New audio models

Added gpt-audio-1.5: OpenAI’s first generally available audio model. It accepts audio inputs and outputs and can be used through the Chat Completions REST API, suited for voice conversations, audio understanding, and audio generation workflows.
Added gpt-4o-transcribe-diarize: an ASR model with built-in speaker diarization that associates audio segments with different speakers in a conversation. This model is available only in the Transcription API.

LLM Router introduction page now live

The LLM Router introduction page is now live, showing public routing dimension scores and the current model pool. Two open endpoints are also available: LLM Router Model Scope returns model scores across 5 categories and 23 sub-dimensions, price coefficients, first-token latency, and the current model pool; Model Vendor Icons returns vendor display names and icons. Image and video generation dimensions are not included yet. See the LLM Router documentation for usage.

Tools and prompt caching

More consistent API error responses

When /v1/responses is called with an unknown model, it now returns HTTP 400, matching other API entrypoints and making the “no available service” case easier for clients to handle instead of receiving a 500.

Request validation failures return earlier

If request-body conversion or parameter validation fails, the API now returns the relevant 400/error response directly instead of forwarding the original request to the model provider, reducing invalid calls and confusing failures.

More stable Claude Code cache hits

Claude Code calls to Claude through Bedrock now have more stable prompt-cache hits within the same session, helping reduce duplicate cache-write cost and first-token latency.

More precise Gemini multi-turn and cache stickiness

Gemini requests now stay on the same service only when the response actually contains thinking signatures or cache activity. This reduces multi-turn signature or cache-state mismatches, while stateless image-generation jobs are no longer unnecessarily pinned.

GPT-5.6: Tools and prompt caching

New models

Added gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna, the three GPT-5.6 family models (officially released by OpenAI on 2026-07-09). All three tiers have a 1,050,000-token context window, 128K max output, and a 2026-02-16 knowledge cutoff, support text and image input, and can be called through Chat Completions, Responses, and the Claude-compatible Messages API. Sol is the flagship tier for complex professional work, described by OpenAI as its best coding model to date; Terra matches GPT-5.5 performance at half the price; Luna targets cost-sensitive workloads.

GPT prompt caching documentation published

Added the GPT Prompt Caching documentation: starting with the GPT-5.6 family, cache writes are billed at 1.25x the input rate, cache reads at 0.1x, and caches are retained for at least 30 minutes; it covers the prompt_cache_key and explicit cache breakpoint parameters, billing logic, API examples, and cache-hit troubleshooting. The OpenAI caching sections of Prompt Caching and Claude Prompt Caching have been updated to match.

Claude Fable/Mythos thinking compatibility

Claude Fable and Mythos requests that use reasoning_effort now route to adaptive thinking, reducing model provider 400 errors for these model families without client-side parameter changes.

Stop sequence support for Claude and Gemini

OpenAI-compatible stop settings are now mapped to Claude and Gemini native requests. Claude-only invalid blank stop sequences are filtered, while OpenAI and Gemini limits are handled per provider, so generation stops more consistently across protocols.

gpt-chat-latest token limit compatibility

gpt-chat-latest and future GPT/ChatGPT latest aliases now preserve max_completion_tokens when required, reducing 400 errors caused by sending the older max_tokens field.

Clearer API Key quota exhaustion message

When a Key-level usage limit is exhausted, the API now returns clearer guidance to adjust and reactivate the Key limit instead of a low-level quota error.

Tool-call response shape aligned with OpenAI

Non-streaming chat responses that contain tool_calls and no text now explicitly return content: null, improving compatibility with OpenAI-style SDKs and response parsers.

Structured output repair and model updates

Structured Output Repair: automatic repair of malformed JSON

Added a Key-level Structured Output Repair capability, disabled by default. Once enabled, for non-streaming requests that declare structured JSON output, when the JSON returned by the model has format errors such as truncation, trailing commas, or code-block wrapping, the gateway automatically repairs it into parseable, valid JSON before returning, keeping numeric values unchanged and requiring no client changes. Supports four protocols: Chat Completions, Responses, Claude, and Gemini; when a repair occurs, the response carries the X-JSON-Repaired: true response header.

Claude prompt caching docs: per-model minimum cache token thresholds completed

Updated the Claude Prompt Caching and Prompt Caching docs with the per-model minimum cacheable token thresholds (512 / 1,024 / 2,048 / 4,096), adding current models such as Claude Opus 4.8, Opus 4.7, and Fable 5. The threshold is not proportional to the model version — a prefix below it is not cached even when marked with cache_control.

Usage dashboard and settlement data alignment

Usage dashboard aggregation is now closer to request-log settlement records by reducing concurrent aggregation loss and aligning the hourly bucket used for delayed replays. Historical dashboard numbers may still have small differences; settlement should continue to use request logs as the source of truth.

New model

Added grok-4.5, a model for coding, agentic tasks, and knowledge work, with configurable reasoning, tool calling, and a 500K context window. It is suited for code repair, complex engineering tasks, knowledge QA, and agent workflows.

GPT / Responses: Tool calling improvements

gpt-5.5+ tool calls automatically bridge to Responses

When using the OpenAI-compatible /v1/chat/completions endpoint with gpt-5.5 or newer models, requests that include tools and reasoning_effort now automatically use Responses capabilities. This reduces model provider 400 failures for that parameter combination without requiring clients to migrate to /v1/responses. See: Responses API

Claude / Vertex AI: Cross-protocol compatibility

Structured output compatibility across protocols

OpenAI-compatible response_format and Claude-native output_config.format are now adapted across the relevant paths. Clients switching between OpenAI and Claude-style protocols can preserve structured output constraints more reliably. See: Structured Output.

More stable `/v1/messages` calls on Vertex AI

When calling Gemini and other non-Claude models through Vertex AI, /v1/messages requests are now routed by model family, reducing not found or 404 failures caused by sending those calls to Claude-specific paths.

Gemini / Tencent Hy3: Models and platform improvements

Gemini native endpoints completed

When using the @google/genai SDK through AIHubMix’s Gemini entrypoint, native embeddings, interactions create, and context caching workflows are now supported. These paths previously could return unregistered-route errors or 404s; they can now be used for embedding generation, interactive calls, and cache create/read/update/delete workflows. See: Gemini Native SDK Integration

Long-context model billing coverage expanded

hy3-preview, ERNIE, Grok, Mimo, and other long-context models now support tiered billing by context length, with more accurate pricing.

Vertex AI streaming log metrics fixed

Vertex AI Gemini/Claude streaming requests now record more accurate time-to-first-token and latency data, making request logs and monitoring more reliable. This does not change model responses.

New model

Added tencent-hy3, a text generation model (Tencent Hunyuan Hy3 official release). MoE architecture with 295B total / 21B active parameters and a 256K context window. Supports hybrid fast-and-slow thinking inference modes, suited for complex reasoning, code generation, and agent workflows. Open-sourced under the Apache 2.0 license.

Jina: Search, file uploads and response reliability

Jina Search/Reader now support POST and file uploads

Jina Search and Reader now support POST requests. Reader can accept local files via multipart upload, including PDF, Word, Excel, PPT, HTML, and images. Default responses are closer to Jina’s native markdown output; clients that need JSON can explicitly send Accept: application/json.

More stable Azure-style service calls through Responses

Fixed an issue where /v1/responses calls bridged to Azure-style non-GPT services could fail with 404 because api-version was empty, reducing direct failures for such requests.

Streaming truncation fixed

Fixed a case where some vLLM / Azure Foundry-style model inference services could randomly truncate streaming responses. Users should more reliably receive complete answers, finish signals, and usage information.

New model: Command A Plus 05-2026

Added command-a-plus-05-2026, a text generation model for chat, content generation, and agent workflows.

Platform improvements

Jina search and webpage reader APIs are now available

Added Jina search and Reader capabilities. You can now use your AIHubMix API Key to fetch Jina search results and read webpage content, suitable for external web lookup, document reading, and Agent tool-use workflows. See: Jina AI

Veo 3.1 image-to-video now supports first/last frames and reference images

Veo 3.1 image-to-video now supports first frame, last frame, and reference image inputs, giving users more precise control over start/end frames, character references, and style references for more advanced video generation workflows. See: Video Generation

Improved Gemini retry stability

Improved compatibility for Gemini requests during retry and fallback scenarios, reducing certain 400 errors and improving fallback success rates. See: Gemini Guides

OpenAI-compatible response fields aligned

OpenAI-compatible non-streaming and streaming responses now preserve null values for fields such as logprobs, refusal, and finish_reason, reducing compatibility differences for SDKs, Agents, and log parsers that expect the standard OpenAI response shape.

Improved Claude Code privacy and compatibility

Improved request compatibility when Claude Code calls Claude models through AIHubMix, reducing client-environment information included in requests sent to the model provider and improving privacy protection for third-party Agent clients.

The self-hosted account system now covers email-code login/registration, password setup, profile editing, third-party account binding/unbinding, and account deletion, with stronger OTP-purpose isolation and disabled-account checks. Stripe top-up callbacks are also handled more reliably.

New models

claude-sonnet-5 for chat, reasoning, and Agent workflows.
gemma-4-31b and longcat-2.0 for text generation.
gemini-3.1-flash-lite-image for image-capable workflows.
mai-image-2.5 and mai-image-2.5-flash for image generation.

Image billing and parameter compatibility

Improved image generation billing and parameter compatibility

gpt-image-2 billing on Images endpoints is now consistently token-based. GLM image generation now passes through extended parameters such as watermark_enabled and quality, so watermark and quality controls can take effect when supported by the model provider. See: Image Generation

Cleaner Gemini file upload logs

Failed Gemini file upload calls no longer create user-visible normal request logs, reducing noise from non-inference endpoints on the logs page while preserving internal troubleshooting logs.

OpenAI Deep Research: New models

New Deep Research models

Added o4-mini-deep-research and o3-deep-research, available only through the /v1/responses endpoint. Requests must include web_search_preview or mcp tools, making them suitable for deep web research and research-style answers.

Responses / Codex: Integration and compatibility

Responses protocol now supports any model

The /v1/responses endpoint is no longer limited to the GPT series and can now call any model on the platform. As a result, tools built on the Responses protocol (such as Codex CLI) can use models like GLM, Gemini, DeepSeek, Kimi, and Qwen through a local model catalog, rather than being restricted to OpenAI’s own models.

Step 3.7 Flash Responses output fix

Fixed an issue where step-3.7-flash could return blank content or an empty response through /v1/responses; reasoning content and final answers now return correctly.

Codex CLI adds custom model support

A “Using Custom Models in Codex” tutorial has been added to the Codex CLI doc: declare any AIHubMix model (GLM, Gemini, DeepSeek, Kimi, Qwen, etc.) via a local model catalog (model_catalog_json) and switch freely from Codex’s /model list, no longer limited to OpenAI’s own models. Includes a one-command script to generate a top-30 catalog and common pitfalls. See: Codex CLI · Using Custom Models in Codex

Inferera: Alternative API domain

Backup domain now supported

Added a backup domain https://api.inferera.com with endpoints and capabilities identical to the primary domain https://aihubmix.com. When the primary domain is unreachable (e.g. connection failures or timeouts), replace the request URL with the backup domain; the API Key, model, request body and other parameters stay the same.

Doubao / HappyHorse: Models and platform improvements

New models

Doubao doubao-seed-2-1-pro, doubao-seed-2-1-turbo.
HappyHorse video series happyhorse-1.1-t2v (text-to-video), happyhorse-1.1-r2v (reference-to-video), happyhorse-1.1-i2v (image-to-video).

Stripe checkout experience improved

Stripe checkout now pre-fills the account email and reduces unnecessary name and billing-address collection, making top-ups with payment methods such as Alipay smoother.

Negative-balance account limit

When an account balance is below -$1, free models can no longer be called until the account is topped up.

CLI, smart routing and agent skills

Added AIHubMix CLI documentation

Added the AIHubMix CLI docs: a single binary with zero dependencies (no Python / Node / Go required) for querying your account balance, managing API Keys, and viewing available models right from the terminal — with script- and AI Agent-friendly output (e.g. Claude Code).

LLM Router

Set model to auto and the gateway analyzes your request to pick the best model from hundreds available, with cost / quality / latency policies, billed by the resolved model, and zero client code changes. See: LLM Router

DeepSeek cache hit rate fix

Fixed an issue where cache hit rates for deepseek-v4-pro and deepseek-v4-flash were lower than expected.

AIHubMix Skill (AI coding assistant extension)

Provides local extension capabilities for AI Agents that support Skills, such as Codex, Claude Code, Cursor, and Cline: complete AIHubMix integration, model lookup, capability-based model selection, example generation, and error troubleshooting in natural language. The Skill reads real-time information such as models, prices, and protocol contracts from the official API on demand, so the Agent does not rely on outdated memory. See: Skills

Kling: Video generation updates

Added Kling video models

Integrated the full Kling video generation capabilities: text-to-video, image-to-video, multi-image reference, and omni multimodal generation, calling the native protocol by model name.

OpenClaw / GLM-5.2: Integration and compatibility

OpenClaw integration plugin issue fixed

AIHubMix’s OpenClaw integration plugin aihubmix-auth has fixed its previous integration issues and is now stable to use. Simply install it and fill in one AIHubMix Key to call OpenAI / Anthropic / Gemini models simultaneously in OpenClaw.

New models

Zhipu glm-5.2.

Open Design: Integration and compatibility

Added Open Design support

AIHubMix is now a built-in BYOK gateway in Open Design (the open-source, local-first Claude Design alternative). Select AIHubMix in its API (BYOK) mode and fill in a single Key to power chat / image / video / speech generation all at once, routing each provider’s native protocol by model name. See: Open Design setup guide

Zhipu GLM 5.2 supports reasoning effort levels

The native Zhipu service’s glm-5.2 supports reasoning_effort to tune thinking depth in tiers; older model versions are selected automatically by version, with no changes required by the caller.

New models

kimi-k2.7-code-highspeed (Kimi code high-speed edition).

Coding GLM-5.2: New models

New models

coding-glm-5.2 and the free edition coding-glm-5.2-free.

Model mapping and fallback

Model mapping and error fallback

Added Model Mapping and error Fallback: Configure per-API-Key model-name mapping and error fallback in the console, rewrite the client’s model alias to the real model name, automatically switch to a backup model when the primary fails, bill by the final responding model, with zero client code changes. See: Model Mapping and Fallback

New models

kimi-k2.7-code.

Step 3.7 Flash / Claude: Models and pricing updates

step-3.7-flash at a limited-time 90% off

step-3.7-flash limited-time offer at 90% off: input is just USD 0.022 per million tokens and output just USD 0.132 per million tokens — give it a try.

Model removal and auto routing

claude-opus-4-20250514 and claude-sonnet-4-20250514 will be officially removed on June 15. At that point, the platform will automatically route the removed models to the 4-5 version of the same series.

Claude Fable 5: Model integration and API limits

New models

claude-fable-5 [Retired].

Claude new models Fable 5 / Mythos 5 feature notes

Fable 5 has stronger safety guardrails, so some normal requests may also be blocked: It applies additional classification to areas such as cybersecurity, biology / chemistry, model distillation, and reasoning extraction. Some technical research, vulnerability analysis, and biomedical questions may be refused or routed to Opus 4.8.
Mythos 5 is restricted access, not a generally available model: Mythos 5 and Fable 5 come from the same capability source, but Mythos 5 uses fewer safety classifiers. It is currently available only to Project Glasswing / approved customers; most users will use the guarded Fable 5.
Higher API cost: Fable 5 is priced at USD 10 per million input tokens and USD 50 per million output tokens, about twice the cost of Opus 4.8.
Privacy: Fable 5 / Mythos 5 are Covered Models, require at least 30 days of data retention by default, and do not support Zero Data Retention.
API refusal behavior: When Fable 5 refuses a request, the API returns HTTP 200, but stop_reason is refusal.

Gemini: Integration and compatibility

Gemini compatible interface supports audio input

When calling Gemini through the OpenAI-compatible interface (/v1/chat/completions), input_audio audio input is now supported (previously it was silently dropped), and audio_tokens metering is now included in the returned usage.

Grok / Hy3 / Step: New models

New models

grok-build-0.1, hy3-preview, and the free model step-3.7-flash-free.

Qwen 3.7 Plus: New models

New models

Tongyi Qianwen qwen3.7-plus.

MiniMax M3 / Baidu: New models

New models

MiniMax minimax-m3.
Baidu musesteamer-air-image (image generation).

Request logs and performance metrics

Log Details Page

Latency: Indicates how quickly a request starts (time from request initiation to the first token returned)
Throughput: Provides a clear measure of the model’s output speed
E2E Latency: The total time from request submission to completion, used to evaluate overall request performance
Provider: Identifies the model vendor that handled the request
Status: Shows the execution result (e.g., success / failure) for quick detection of anomalies
TID: A unique request identifier that can be shared with support for faster troubleshooting

Global acceleration and health monitoring

AIHubMix Global Acceleration Network is now live: Built on self-managed global edge nodes and an intelligent routing system, with continuous monitoring and dynamic optimization, achieving 75% lower latency, 60% improved stability, and 99.99% availability—delivering a faster, more reliable AI experience.
Added 24/7 real-time health monitoring: Distributed probes scan the entire network every minute, tracking latency, success rate, and stability. Issues are detected and resolved before users even notice, ensuring consistent performance.
Enhanced intelligent traffic routing: Node health is dynamically evaluated across multiple time windows, enabling millisecond-level switching to the optimal path, significantly reducing fluctuations and timeouts while improving overall request success rates.

Chat / Responses: Cross-protocol compatibility

New Feature: Chat → Responses Compatibility
This release introduces Chat → Responses compatibility, allowing the Chat Completions API to invoke OpenAI models that support the Responses protocol only, including gpt-5.2-codex, gpt-5.1-codex-max, and gpt-5.2-pro. If you want to force the AIHubmix Chat API to route requests through the Responses protocol, add the following header to your request:
X-Use-Responses-Enabled: true When a model supports both Chat and Responses protocols, setting this header will force the request to use the Responses API.
Please note that the Responses protocol does not currently support audio input or output, so plan usage accordingly.
Model Deprecation Notice:
OpenAI will deprecate chatgpt-4o-latest on February 17, 2026. After the deprecation, we will automatically remap chatgpt-4o-latest to gpt-4o-2024-11-20.

2025

Google Files API: File upload support

New feature: Google API now supports Files API.

Qwen / Seedream / Baidu: Models and media improvements

Added support for Qwen series, Doubao Seedream 4, and Baidu image generation models

AIHubMix: Models and MCP image tools

Released the Aihubmix Image Generation MCP, making it easier for developers to integrate image generation services

Claude Code: Integration and compatibility

Use any large language model on the AiHubMix platform directly in Claude Code

AI SDK: Integration and compatibility

Added support for the AI SDK: access a massive number of models with a single API key

Flux: Models and media improvements

Added support for the Flux image generation API, enabling high-quality images in seconds

Qwen Code: Integration and compatibility

Added support for Qwen Code, leveraging all large language models available on the Aihubmix platform

llms.txt: Documentation updates

Added support for llms.txt: get standardized model navigation with one click so your LLM assistant can quickly understand the entire model ecosystem

Gemini CLI / Responses: Integration and compatibility

Added forwarding support for Gemini CLI, with multiple flexible usage modes
Added code interpreter and Remote MCP invocation to the OpenAI Responses API

Unified image generation API

Added a Unified Image Generation API, supporting major models including OpenAI, Ideogram, Stability, and Google Imagen

APP-Code: Integration and compatibility

Launched APP-Code, offering developers a 10% discount across all models

HTTP status code documentation

Added HTTP Status Code documentation to help users better understand error messages

Veo 3.0: Video generation updates

Added support for Veo 3.0 video generation to expand creative formats

Responses: Reasoning summaries

Added support for OpenAI Reasoning Summaries in the Responses API

Gemini: Prompt caching support

Added implicit caching for Gemini, with automatic cache hits and hit feedback
Developers can use usage_metadata to determine cache hits
Cost savings are not guaranteed and depend on request structure and usage scenarios

Claude 4: Tools and prompt caching

Full Support for New Claude 4 Features

⏳ New cache TTL: 1-hour cache support ^Beta
🎉 New text editing tools: Claude 4 now supports text_editor_20250429 and str_replace_based_edit_tool
🚫 New refusal stop reason for safety-based rejections
🧠 Extended Thinking: Claude 4 now returns full summaries of its reasoning process
🔄 Interleaved Thinking: Tool usage can now interleave with extended thinking for more natural conversations (Beta)
⚠️ Deprecated Features:
- undo_edit is no longer supported
- token-efficient-tools-2025-02-19 removed (Claude 3.7 only)
- output-128k-2025-02-19 removed (Claude 3.7 only)
📚 Full migration guides and code examples have been updated to help users smoothly transition from Claude 3.7 to Claude 4

Dify: Integration and compatibility

Added support for the Dify plugin, enabling seamless integration of Aihubmix models into Dify
Extend and manage over 200 models with a single API key

Codex Mini / Google: Models and media improvements

Added support for codex-mini-latest, optimized for programming tasks, accessible via the Responses API or Codex CLI
Added support for Google Imagen 3.0 image generation and Veo 2.0 video generation
gemini-2.0-flash-exp upgraded to the official preview version gemini-2.0-flash-preview-image-generation

Ideogram V3: Models and media improvements

Added the Ideogram AI V3 API — Ideogram’s most advanced image generation model

CLI: Integration and compatibility

Added Utility Management Scripts for managing API keys, viewing accounts, and listing available models via CLI

GPT Image 1 / Gemini: Models and media improvements

The highly anticipated OpenAI image generation API gpt-image-1 is now live, supporting both text-to-image and image-to-image
Added native Gemini API support with precise reasoning budget control for Flash 2.5

Jina: Search integration

Integrated three core Jina AI APIs to help build powerful agents: Embeddings, Rerank, and DeepSearch

OpenAI Responses API: Integration and compatibility

Added support for the OpenAI Responses API endpoint with expanded tool capabilities

Codex CLI: Integration and compatibility

Added OpenAI CodeX CLI support: program with natural language directly in the terminal

Search integration

By appending :surfing to a model ID, any model can gain search capabilities (Beta)

Claude: Prompt caching support

Added Claude prompt caching, saving up to 76% in cost for repeated high-frequency prompts

Ideogram: Models and media improvements

Added Ideogram AI image generation support with strong text rendering, hybrid generation, local edits, and upscaling

Documentation updates

Brand-new documentation experience launched

Claude: Text editing tools

Added support for the Claude Text Edit Tool

Brand identity update

Launched the brand-new Trident logo

OpenAI / Gemini: Search integration

Added native search support for OpenAI and Google Gemini models
Third-party search integration will be added in future updates

GPT-4o Search: New models

Added models: gpt-4o-mini-search-preview and gpt-4o-search-preview

o1 / o3-mini: Pricing adjustments

Prices for o1 and o3-mini reduced by 10%, in line with official pricing

DeepSeek / Qwen: Models and pricing updates

Due to a 7× price increase from Microsoft, the price of aihubmix-DeepSeek-R1 also increased 7×
Recommended alternative: Volcano Engine’s DeepSeek-R1 (more stable and cost-effective)
Added models: qwen-qwq-32b and qwen2.5-vl-72b-instruct

Claude / GPT-4.5 Preview: Models and pricing updates

All Claude models received a 15% price reduction
Added model gpt-4.5-preview (extremely expensive — use with caution)

DeepSeek: Model service updates

Improved DeepSeek stability
ByteDance versions of DeepSeek are currently the most stable
Recommended models: DeepSeek-R1 and DeepSeek-V3

Claude 3.7 Sonnet: New models

Added model claude-3-7-sonnet-20250219

Model service updates

The gpt-4o model may occasionally respond very slowly due to model provider issues
It is recommended to temporarily switch to gpt-4o-2024-11-20
The Perplexity API is temporarily offline
Due to Perplexity’s complex billing model and higher costs than this platform’s pricing structure, the service will be relaunched after pricing adjustments
The temporary ByteDance official discount has ended and prices have returned to normal
The price of DeepSeek-R1 has been increased accordingly
Added a new model details page with full parameter information

DeepSeek-V3: Pricing adjustments

The temporary ByteDance official discount has ended and prices have returned to normal
The price of DeepSeek-V3 has been increased
ByteDance’s R1 model is also expected to return to normal pricing soon, and this platform will同步 adjust prices accordingly

Models and platform improvements

Added model: kimi-latest
(Official billing is tiered by input length at 8k, 32k, and 128k.
This platform does not support tiered pricing and uses the mid-tier 32k as the pricing standard.
If you are price-sensitive, use with caution.)
Optimized overall website layout
Merged the Changelog page into the Usage Statistics page
Moved Announcements to the Model Marketplace page
Moved Settings under the user avatar menu
Reduced the price of aihubmix-DeepSeek-R1 by 50%
Added models:
gemini-2.0-pro-exp-02-05-search, gemini-2.0-flash-exp-search
(Integrated with Google’s official online search)
Added models:
gemini-2.0-flash, gemini-2.0-pro-exp-02-05, gemini-2.0-flash-lite-preview-02-05
Added models:
o3-mini, o1
(These two models are billed about 10% higher than official pricing due to limited account resources)

o1 / o3-mini: Model integration and API limits

The o1 model does not support the stream parameter in the official OpenAI API
The o3-mini model does not support the temperature parameter
A new parameter reasoning_effort is available with values: "low", "medium", "high"
Default is "medium" if not specified

OpenAI: Audio generation and model updates

Feature Update:

Added support for OpenAI audio model input and output
The api.aihubmix.com preview server is now available
After one week of stable operation, the main site will be updated
Backend billing is fully consistent with official pricing
Currently, usage logs only display text token usage
Audio token usage is not yet shown in logs but does not affect normal usage

New Models Added:

o3-mini, o1
(Billed about 10% higher than official pricing due to limited account availability)
aihubmix-DeepSeek-R1 (recommended, highly stable)
qwen-max-0125 (Qwen2.5-Max), sonar-reasoning
deepseek-ai/DeepSeek-R1-Zero, deepseek-ai/DeepSeek-R1, deepseek-r1-distill-llama-70b
aihub-Phi-4
Doubao-1.5-pro-256k, Doubao-1.5-pro-32k,
Doubao-1.5-lite-32k, Doubao-1.5-vision-pro-32k
sonar, sonar-pro (latest from Perplexity AI)
gemini-2.0-flash-thinking-exp-01-21
deepseek-reasoner (aka DeepSeek-R1)
MiniMax-Text-01
codestral-latest (Mistral’s new code model — Codestral 25.01)

DeepSeek / Doubao / Gemini: New models

New Models Added:

aihub-Phi-4
Doubao-1.5-pro-256k, Doubao-1.5-pro-32k,
Doubao-1.5-lite-32k, Doubao-1.5-vision-pro-32k
sonar, sonar-pro (latest from Perplexity AI)
gemini-2.0-flash-thinking-exp-01-21
deepseek-reasoner (aka DeepSeek-R1)

Perplexity / MiniMax / Codestral: New models

Added Perplexity AI API models
Currently supported only on the preview server api.aihubmix.com
After stable testing, it will be rolled out to the main server aihubmix.com
api.aihubmix.com is the preview server
New features will be deployed there first and promoted to the main server after ~1 week of stability testing

New Models Added:

MiniMax-Text-01
codestral-latest (Mistral Codestral 25.01)

Gemini / DeepSeek-V3: New models

Added gemini-2.0-flash-exp-search, supporting native Google online search
The official Gemini 2.0 Flash model requires additional parameters for online search
Aihubmix has integrated this functionality — simply append search to the model name
Added model: deepseek-ai/DeepSeek-V3

Model discovery and details

Launched the new Model Marketplace page to replace the old Model & Pricing page

2024

Gemini: Model service updates

Fixed the issue where gemini-2.0-flash-thinking-exp-1219 only returned reasoning without final answers
Fixed the issue of balance reminder emails not being delivered

Usage statistics and new models

Added Usage Statistics page
Added Recharge History page
Added Doubao series models:
Doubao-lite-128k, Doubao-lite-32k, Doubao-lite-4k,
Doubao-pro-128k, Doubao-pro-256k, Doubao-pro-32k, Doubao-pro-4k
Added model: gemini-2.0-flash-thinking-exp-1219
Added models:
gemini-2.0-flash-exp, aihubmix-Mistral-Large-2411,
aihubmix-Llama-3-3-70B-Instruct, grok-2-1212, grok-2-vision-1212
Added models:
gemini-exp-1206, llama-3.3-70b-versatile, learnlm-1.5-pro-experimental

Gemini / Mistral / Llama: New models

Added models:
gemini-2.0-flash-exp, aihubmix-Mistral-Large-2411,
aihubmix-Llama-3-3-70B-Instruct

Usage statistics and new models

Added models:
gemini-exp-1206, llama-3.3-70b-versatile, learnlm-1.5-pro-experimental
Added Usage Statistics page

GPT-4o / Grok / Qwen: New models

Recently added models:
gpt-4o-2024-11-20, step-2-16k, grok-vision-beta
Qwen 2.5 Turbo million-context model:
qwen-turbo-2024-11-01

Claude SDK / Messages: Integration and compatibility

Added compatibility with the native Claude SDK
The v1/messages endpoint is now live
Native Claude prompt caching and computer use features are not yet supported
These will be completed within the next two weeks

Claude Haiku / Grok Beta: New models

Added model: claude-3-5-haiku-20241022
Added Elon Musk’s xAI latest model: grok-beta

Claude 3.5 Sonnet: New models

Added model: claude-3-5-sonnet-20241022

OpenAI: Prompt caching support

OpenAI’s latest caching feature is now live
Currently supported models:
- GPT-4o
- GPT-4o-mini
- o1-preview
- o1-mini
Note: gpt-4o-2024-05-13 is not included in the official supported list
Cache hit tokens will be visible in backend logs when a request hits the cache
For full details and usage rules, refer to the OpenAI official documentation

GPT-4o / Llama / Cohere: Models and pricing updates

Backend billing for gpt-4o has been reduced to match official pricing
Added models:
aihubmix-Llama-3-2-90B-Vision, aihubmix-Llama-3-70B-Instruct
Added Cohere latest models:
aihubmix-command-r-08-2024, aihubmix-command-r-plus-08-2024

Whisper: Audio models and billing

Added models: whisper-large-v3 and distil-whisper-large-v3-en
Note: Whisper model billing is based on input seconds
The current pricing display on the site is incorrect and will be fixed
Backend billing for whisper-1 fully matches OpenAI official pricing

o1: Model integration and API limits

Added models: o1-mini and o1-preview
Note: These models require updated parameters
Some client shells may throw errors if defaults are not updated

Test results show that the o1 model does NOT support:

system field → 400 error
tools field → 400 error
Image input → 400 error
json_object output → 500 error
structured output → 400 error
logprobs output → 403 error
stream output → 400 error

Rate limits and fixed parameters:

o1 series: 20 RPM, 150,000,000 TPM — extremely low, frequent 429 errors possible
temperature, top_p, and n are fixed at 1
presence_penalty and frequency_penalty are fixed at 0

Model service updates

Added model: mattshumer/Reflection-Llama-3.1-70B
(Reported to be one of the strongest fine-tuned versions of LLaMA 3.1 70B)
Claude-3 model prices increased
To ensure stable supply, calls through this platform are currently ~10% more expensive than direct official usage
Increased concurrency capacity for OpenAI models
The system now theoretically supports near-unlimited concurrency

Phi / Llama / Claude: Models and platform improvements

Added models:
Phi3medium128k, ahm-Phi-3-medium-4k, ahm-Phi-3-small-128k
Improved stability for LLaMA-related models
Further optimized compatibility for Claude models

GPT-4o / Gemini: New models

Added OpenAI’s newly released gpt-4o-2024-08-06
See: https://platform.openai.com/docs/guides/structured-outputs
Added Google’s latest model: gemini-1.5-pro-exp-0801

Online payments and API compatibility

Added direct online payment for account top-ups
Fixed Claude multi-turn conversation format error:
messages: roles must alternate between "user" and "assistant", but found multiple "user" roles in a row
Optimized index handling when using function calling with Claude models
The backup server https://orisound.cn will be fully decommissioned on Sep 7
Please migrate to the main server https://aihubmix.com or backup server https://api.aihubmix.com

Mistral Large 2: Models and platform improvements

Added support for Mistral Large 2
Model name: Mistral-large-2407 or aihubmix-Mistral-large-2407
System optimizations

Llama 3.1: New models

Added latest LLaMA 3.1 models:
llama-3.1-405b-instruct, llama-3.1-70b-versatile, llama-3.1-8b-instant

GPT-4o Mini: Image billing and parameter compatibility

Fixed pricing calculation issues for the gpt-4o-mini model
- Text input pricing: 1/33 of GPT-4o officially
- Image input pricing: equal to GPT-4o
To align with official pricing, image token counts for gpt-4o-mini are multiplied by 33 during billing
Refer to OpenAI official pricing for details

GPT-4o Mini: New models

Added support for the gpt-4o-mini model
Backend billing is fully aligned with official pricing

Streaming usage reporting

Added support for the official include_usage API parameter
This allows returning usage data in stream mode
See the official documentation for details

Integration and compatibility

The new version of NextWeb now supports calling non-OpenAI models through this platform
Added backend billing support for Alibaba Qwen models
Calls through this platform cost ~10% more than direct Alibaba Cloud usage
Improved Azure OpenAI output compatibility with the standard OpenAI API
Added tool calling support for Claude-3
Added many new models (see Settings → Available Models)

Request logs and performance metrics

Overall backend UI optimization
Each log entry now displays the model unit price at the time of the request
Added the Model & Pricing page

Claude 3.5 Sonnet: New models

The latest claude-3-5-sonnet-20240620 is now supported
See the guide for calling non-OpenAI models on this platform

Request history downloads

Backend logs now support downloading historical request records

Azure OpenAI: Model routing and API updates

The probability of randomly routing requests to Azure OpenAI has been significantly reduced

Claude 3: Pricing adjustments

Reduced backend costs for Claude-3 models
(Claude 3 Haiku, Claude 3 Sonnet, Claude 3 Opus)
Backend billing now matches official pricing
As a result, the effective retail API cost on this site is equivalent to ~86% of official pricing

GPT-4o: Token counting improvements

Optimized GPT-4o token billing
Tokenizer changed from cl100k_base to o200k_base
As a result, streaming token counts for Chinese, Korean, and Japanese are lower than before

Qwen2: New models

Added Alibaba’s latest open-source Qinwen 2 models:
- alibaba/Qwen2-7B-Instruct
- alibaba/Qwen2-57B-A14B-Instruct
- alibaba/Qwen2-72B-Instruct

GPT-4o / Gemini / Llama: New models

Added model: gemini-1.5-flash
Added model: gpt-4o
Added models:
llama3-70b-8192, llama3-8b-8192,
gemini-1.5-pro, command-r, command-r-plus
Claude-3 model supply has been restored
Endpoints are currently deployed across AWS and Google Cloud
To cover infrastructure and operational costs, Claude-3 backend billing is ~10% higher than official pricing
With increased usage, this will be gradually reduced to ~5% or lower
Concurrency limits are currently under testing and will be increased as demand grows

Last updated: 2026-06-22

​New models

​New models

​New models

​Image input for Realtime conversations

​2026September 23

​2026September 22

​2026September 20

​2026September 18

​2026September 17

​2026

​Realtime speech transcription is live (gpt-live-transcribe)

​New models

​New feature: realtime speech transcription

​Gemini streaming tool-call compatibility is improved

​Model capability metadata is more complete

​User notification email delivery is steadier

​Media compatibility updates

​Multi-turn reasoning compatibility is improved

​MAI image editing is steadier

​Media failure messages are more actionable

​Responses image-tool billing and Azure deployments are steadier

​Image-tool calls are more stable

​Image consumption records are easier to reconcile

​Qwen 3.8 27B: New models

​New models

​GPT Image 2.5: Models and media improvements

​New models

​Image billing fallback is more reliable

​Public model information is more consistent

​Media errors are clearer and more consistent

​Playground endpoint filtering is more accurate

​Legacy Claude system-role requests are more compatible

​DeepSeek V4.1 Flash: Models and platform improvements

​New models

​Gemini billing is safer when usage is missing

​Model routing errors are easier to diagnose

​GPT and Claude tool-call compatibility is steadier

​Temporary file upload API added

​GPT: Tools and prompt caching

​Model badges are more consistent

​GPT tool-call compatibility is improved

​GPT cache-write billing tiers are more accurate

​GPT-6 Astra / MAI Image 2.6: New models

​New models

​Image generation reliability

​Image generation requests are more stable

​Media task errors are more consistent

​Long-running agents and multimodal models

​New models

​Claude Fable / Qwen / MiniMax: New models

​New models

​Model availability and media input

​Retired-model errors are clearer

​Media input errors are more actionable

​Google image input validation is temporarily higher

​Hy4 Preview: New models

​New models

​Qwen 3.8 Flash: Models and MCP image tools

​New models

​MCP image-tool calls are clearer

​Broadcast, remote MCP and Access Key

​New models

​New feature

​MCP and Access Key entry points are more complete

​Transparent-background image requests are steadier

​Agent-facing files update more quickly

​API migration guidance

​Legacy media API migration errors are clearer

​Wan 3.0 / GPT-5.6 Sol: Models and pricing updates

​New models

​GPT-5.6 Sol price reduction

​Async media task enablement is clearer

​DeepSeek V4: Models and media improvements

​New models

​Image generation failure recovery is steadier

​Agent integration and model discovery

​Platform announcement

​Image generation API parameters and errors are more consistent

​Model Hub now supports main models with versions

​Async image tasks can move faster

New models

New models

New models

Image input for Realtime conversations

2026

September 23

2026

September 22

2026

September 20

2026

September 18

2026

September 17

2026

Realtime speech transcription is live (gpt-live-transcribe)

New models

New feature: realtime speech transcription

Gemini streaming tool-call compatibility is improved

Model capability metadata is more complete

User notification email delivery is steadier

Media compatibility updates

Multi-turn reasoning compatibility is improved

MAI image editing is steadier

Media failure messages are more actionable

Responses image-tool billing and Azure deployments are steadier

Image-tool calls are more stable

Image consumption records are easier to reconcile

Qwen 3.8 27B: New models

New models

GPT Image 2.5: Models and media improvements

New models

Image billing fallback is more reliable

Public model information is more consistent

Media errors are clearer and more consistent

Playground endpoint filtering is more accurate

Legacy Claude system-role requests are more compatible

DeepSeek V4.1 Flash: Models and platform improvements

New models

Gemini billing is safer when usage is missing

Model routing errors are easier to diagnose

GPT and Claude tool-call compatibility is steadier

Temporary file upload API added

GPT: Tools and prompt caching

Model badges are more consistent

GPT tool-call compatibility is improved

GPT cache-write billing tiers are more accurate

GPT-6 Astra / MAI Image 2.6: New models

New models

Image generation reliability

Image generation requests are more stable

Media task errors are more consistent

Long-running agents and multimodal models

New models

Claude Fable / Qwen / MiniMax: New models

New models

Model availability and media input

Retired-model errors are clearer

Media input errors are more actionable

Google image input validation is temporarily higher

Hy4 Preview: New models

New models

Qwen 3.8 Flash: Models and MCP image tools

New models

MCP image-tool calls are clearer

Broadcast, remote MCP and Access Key

New models

New feature

MCP and Access Key entry points are more complete

Transparent-background image requests are steadier

Agent-facing files update more quickly

API migration guidance

Legacy media API migration errors are clearer

Wan 3.0 / GPT-5.6 Sol: Models and pricing updates

New models

GPT-5.6 Sol price reduction

Async media task enablement is clearer

DeepSeek V4: Models and media improvements

New models

Image generation failure recovery is steadier

Agent integration and model discovery

Platform announcement

Image generation API parameters and errors are more consistent

Model Hub now supports main models with versions

Async image tasks can move faster