Changelog

Integrated with 3 core APIs from Jina AI, empowering you to effortlessly build powerful intelligent agents. The three APIs are: Vector Embeddings, Rerank, and Deep Search.

GPT 4o Image Reverse Interface Early Access.

OpenAI Responses API is supported, offering more comprehensive tool integration.

Added OpenAI CodeX Cli support, Tell Codex CLI what to build, fix, or explain, then watch it bring your ideas to life.

Added :surfing to the model id to enable search functionality for any model ^Beta

Support Claude prompt caching, high-frequency prompts can save up to 76% of costs!

Added support for Ideogram AI drawing interface, featuring powerful text-to-image capabilities, including image generation, remix, partial editing and upscaling.

New docs published, bring you a better experience

Added support for Claude Text Edit Tool

Enabled new Trident Logo

Added native search functionality support for OpenAI and Google Gemini models; future plans include extending this interface to support third-party search

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

o1 and o3-mini prices reduced by 10%, now aligned with official pricing

aihubmix-DeepSeek-R1 price increased 7x due to Microsoft price increase; recommend using Volcano’s DeepSeek-R1 for stability and affordability. Added new models: qwen-qwq-32b and qwen2.5-vl-72b-instruct

Claude models price reduced by 15%; Added new model gpt-4.5-preview (Note: extremely expensive, please use with caution)

Improved Deepseek stability; Most stable from ByteDance, recommended models: DeepSeek-R1 and DeepSeek-V3

Added model claude-3-7-sonnet-20250219

• gpt-4o occasionally experiencing slow responses due to vendor issues; temporarily recommend using gpt-4o-2024-11-20
• Perplexity API temporarily offline due to complex official pricing structure exceeding our platform’s pricing structure; will relaunch after price adjustment
• ByteDance official limited-time discount ended, DeepSeek-R1 price has increased
• Added model details page and parameter information

• ByteDance official limited-time discount ended, DeepSeek-V3 price has increased; expecting R1 price to increase soon, we will adjust accordingly

• Added new model: kimi-latest (Official pricing based on input length tiers 8k/32k/128k, our platform uses 32k tier as standard, price-sensitive users please note)
• Optimized website structure
• Merged logs page into usage statistics page
• Moved announcements to model marketplace page
• Moved settings under profile picture
• aihubmix-DeepSeek-R1 price reduced by 50%
• Added new models: gemini-2.0-pro-exp-02-05-search, gemini-2.0-flash-exp-search with integrated Google official search functionality
• Added new models: gemini-2.0-flash, gemini-2.0-pro-exp-02-05, gemini-2.0-flash-lite-preview-02-05
• Added new models: o3-mini, o1 (Note: backend billing ~10% higher than official due to limited accounts)

• o1 model: OpenAI official does not support stream parameter
• o3-mini does not support temperature parameter; added new parameter “Reasoning effort” accepting “low, medium, high” (defaults to medium)

Feature: Added support for OpenAI voice model input/output functionality. Available on api.aihubmix.com server, main server support coming after 1 week of stability. Overall backend billing matches official rates. Currently, logs only show text tokens; voice costs not yet displayed but functionality unaffected.

New models added:

• o3-mini, o1 (Note: backend billing ~10% higher than official due to limited accounts)
• aihubmix-DeepSeek-R1 (Recommended, more stable)
• qwen-max-0125 (Qwen2.5-Max), sonar-reasoning
• deepseek-ai/DeepSeek-R1-Zero and 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)

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)

• Added Perplexity AI API models; currently only available on preview server api.aihubmix.com, will update to main server aihubmi.com if stable
• api.aihubmix.com is preview server, new features update here first, typically moving to main server aihubmix.com after 1 week of stability

New models added:

• MiniMax-Text-01
• codestral-latest (Mistral’s new code model - Codestral 25.01)
• gpt-4o-zh: automatically translates any language input to English for the model, model output automatically translated back to Chinese; feature in testing phase, limited concurrent requests, only supports gpt-4o model

• Added gemini-2.0-flash-exp-search with Google native web search functionality; official gemini 2.0 flash model’s web search requires additional parameters, aihubmix integrated this - just add ‘search’ to model name to use
• Added new model deepseek-ai/DeepSeek-V3

• Added model marketplace page replacing original model/pricing page

• Fixed gemini-2.0-flash-thinking-exp-1219 model only outputting thoughts without answers
• Fixed issue with balance reminder emails not being received

• Added usage statistics page and recharge record 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 new models: gemini-2.0-flash-thinking-exp-1219, gemini-2.0-flash-exp, aihubmix-Mistral-Large-2411, aihubmix-Llama-3-3-70B-Instruct, grok-2-1212, grok-2-vision-1212
• Added new models: gemini-exp-1206, llama-3.3-70b-versatile, learnlm-1.5-pro-experimental

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

• Added new models: gemini-exp-1206, llama-3.3-70b-versatile, learnlm-1.5-pro-experimental
• Added usage statistics page

• 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

• Compatible with Claude native SDK, v1/messages interface now supported
• Cache and computer usage functions for Claude native interface not yet supported (prompt caching and computer use); will continue improving these features in next two weeks

• Added new model: claude-3-5-haiku-20241022
• Added Elon Musk’s x.ai latest model grok-beta

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

OpenAI’s latest caching feature now online. Currently supports:

• GPT-4o
• GPT-4o-mini
• o1-preview
• o1-mini

Note: gpt-4o-2024-05-13 version not officially supported.

If request hits cache, you’ll see relevant cache token data in backend logs.

For detailed information and usage rules, visit: OpenAI Caching Feature Details

• Reduced backend billing for gpt-4o model, synchronized with official pricing
• Added new models: aihubmix-Llama-3-2-90B-Vision, aihubmix-Llama-3-70B-Instruct
• Added Cohere’s latest models: aihubmix-command-r-08-2024, aihubmix-command-r-plus-08-2024

• Added new models: whisper-large-v3 and distil-whisper-large-v3-en
• Note: Whisper model billing based on input seconds; current page pricing display issue to be fixed. Backend billing accurate, fully synchronized with OpenAI official charges

• Added models o1-mini and o1-preview Note: These latest models require input parameter changes. Some wrapper software may report errors if default parameters not updated.

Important Notes: Testing shows o1 models do 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
• o1 series: 20 RPM, 150,000,000 TPM, very low, 429 errors may occur anytime
• Others: temperature, top_p, and n fixed at 1; presence_penalty and frequency_penalty fixed at 0

• Added new model: mattshumer/Reflection-Llama-3.1-70B (reportedly strongest fine-tuned version of llama3.1-70b)
• Increased Claude-3 model pricing to maintain stable supply, currently 10% more expensive than direct official API use; prices will gradually decrease
• Enhanced concurrent capabilities of OpenAI series models, theoretically supporting unlimited concurrency

• Added new models: Phi3medium128k, ahm-Phi-3-medium-4k, ahm-Phi-3-small-128k
• Improved stability of Llama-related models
• Further optimized compatibility of Claude models

• Added OpenAI’s newly updated 4o version 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

• Added online direct payment recharge
• 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 issue when using function features of Claude models
• Backup server https://orisound.cn will be fully decommissioned on September 7; please switch to main server https://aihubmix.com or backup server https://api.aihubmix.com

• Added support for Mistral Large 2, model name: Mistral-large-2407 or aihubmix-Mistral-large-2407
• System optimization

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

• Fixed pricing calculation issue for gpt-4o-mini model: Text input price: Official gpt-4o-mini model text input price is 1/33 of gpt-4o model price Image input price: Official gpt-4o-mini model image input price equals gpt-4o model price
• For accurate price calculation, we multiply token count for gpt-4o-mini model image input by 33x to align with official price
• More details at OpenAI Official Pricing

• Added support for gpt-4o-mini model, backend billing synchronized with official

• Supports official API parameter include_usage, passing parameters can return usage in stream mode, see Official Documentation

• New nextweb version supports calling non-OpenAI models Call Non-OpenAI Models on Our Site
• Added backend billing for Alibaba Qwen models, overall cost ~10% higher than Alibaba Cloud official
• Optimized Azure OpenAI output compatibility with OpenAI interface
• Supports Claude-3’s tool calling
• Added many new models, see Settings for available models

• Overall backend interface optimized
• Each log request record now shows model unit price at time of request
• Added model and pricing page Model/Pricing

• Latest claude-3-5-sonnet-20240620 now supported, see usage instructions at How to Call Non-OpenAI Models on Our Site

• Backend log page now supports downloading usage request records

• Reduced probability of randomly hitting Azure OpenAI, now almost negligible

• Lowered fees for Claude-3 related models (Claude 3 Haiku, Claude 3 Sonnet, Claude 3 Opus) to align with official backend charges; current retail price ~14% off official price

• Overall service architecture upgrade, all servers and data migrated to Microsoft Azure
• Future plans include secondary deep development and optimization based on open-source version of one API (commercial license obtained through sponsorship)
• Log data too large (over 100 million request logs) for temporary migration; contact customer service for previous log queries
• Optimized token billing for gpt-4o, tokenizer changed from cI100k_base to 0200k_base; results in decreased token count for streaming requests in Chinese, Korean, and Japanese

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

• Added new model gemini-1.5-flash
• Added new model gpt-4o
• Error accessing recharge page in Jiangsu region due to telecom hijacking of recharge domain, please contact customer service for recharge
• Added llama3 (llama3-70b-8192, llama3-8b-8192) gemini-1.5-pro, command-r, command-r-plus
• Claude-3 model supply restored; currently connecting to Claude-3 endpoints deployed on AWS and Google Cloud
• Claude-3 model and price backend billing 10% more expensive than official to cover server costs and team expenses
• Prices will gradually decrease to ~5% or lower as call volume increases
• Current concurrency needs testing; will apply for higher concurrent calls as usage increases