July 2026 Release Spotlight: ~30 New Models and Media APIs

Jul 31, 2026 · AIHubMix · 6 min read · Changelog

July 2026 Release Spotlight: ~30 New Models and Media APIs

This month AIHubMix added nearly 30 new models across chat, coding, audio, image and 3D, and launched several platform capabilities including media generation APIs, structured output repair and a 3D generation API. The range of modalities available through a single API key continues to expand. Here are the highlights.

Media generation APIs and unified async tasks

New media generation endpoints /ai/v1/images/generations and /ai/v1/images/edits are now available, together with unified async tasks, result queries, artifact downloads and webhook callbacks. Image and video workflows can integrate through the same task lifecycle, making it easy to track long-running generation results. Generated artifacts support direct download and batch retrieval. See Async Tasks.

Structured output repair

A new key-level structured output repair capability is available, disabled by default. Once enabled, for non-streaming requests that declare structured JSON output, when the model returns JSON with format errors such as truncation, trailing commas or code block wrappers, the platform automatically repairs it into parseable valid JSON before returning. Values remain unchanged and no client changes are required. It supports the Chat Completions, Responses, Claude and Gemini protocols, and repaired responses carry the X-JSON-Repaired: true header. See Structured Output Repair.

3D generation API

The new /v1/3d/generations async 3D generation API launches with Tencent Hunyuan 3D (hy-3d-3.1) as the first supported model. It covers text-to-3D, image-to-3D and multi-view input, with artifacts in obj, glb, stl, usdz and fbx formats plus PBR materials, suitable for games, e-commerce display, 3D printing and product design. See 3D Generation API.

Official AIHubMix MCP Server

The official hosted entry point for MCP clients is now live: mcp.aihubmix.com/mcp (backup: mcp.inferera.com/mcp). Connect it in Claude Code, Codex, Cursor and other tools to query models and prices, search documentation, check your balance, and call chat, image generation and video generation tools. Credentials are passed per request by the client and the server does not store API keys.

Jina search and web reading

New Jina Search and Reader capabilities let you retrieve search results and web page content with your AIHubMix API key, suitable for external information retrieval, document reading and agent tool calling. Both support POST requests, and Reader also accepts multipart uploads of local files such as PDF, Word, Excel, PPT, HTML and images. See Jina AI.

Other feature updates

  • Qwen TTS audio generation: qwen-audio-3.0-tts-plus and qwen-audio-3.0-tts-flash can be called through /v1/audio/speech. Non-streaming requests return an audio result link, streaming requests return SSE audio generation events, with support for emotion tags and natural language voice instructions. See TTS.
  • GPT prompt caching documentation: starting with the GPT-5.6 series, cache writes are billed at 1.25x the input price, cache reads at 0.1x, and caches are retained for at least 30 minutes. Includes prompt_cache_key parameter details and cache hit troubleshooting. See GPT Prompt Caching.
  • Veo 3.1 image-to-video supports first/last frames and reference images: control the opening and closing shots, character references and style references with more precision. See Video Generation.
  • Automatic Responses bridging for gpt-5.5 and above tool calls: requests sent through /v1/chat/completions with tools and reasoning_effort automatically use Responses capabilities, giving existing clients more reliable tool calling without any changes. See Responses API.
  • Gemini native endpoints completed: the @google/genai SDK now supports native Embeddings, Interactions and Context Caching when called through AIHubMix. See Gemini Native SDK.
  • Streaming connections extended to 2 hours: the maximum streaming connection time has been extended from 1 hour to 2 hours, so long thinking and long generation tasks are less likely to be cut off early.
  • Gemini image parameter passthrough: when calling Gemini image output models through the OpenAI-compatible API, aspectRatio and imageSize can be passed in extra_body.generationConfig.imageConfig.
  • Structured output across protocols: the OpenAI-compatible response_format and Claude native output_config.format now adapt in both directions on the relevant paths. See Structured Output.

Stability and fixes

  • Multi-protocol compatibility improvements: Messages response fields, thinking parameters, stop parameters and tool call response structures are further aligned with each protocol ecosystem.
  • Streaming reliability improvements: fixed truncated streaming output and responses misclassified as empty for some models, with more complete usage records when clients disconnect.
  • Cache metering and billing accuracy improvements, so billing details better reflect actual usage.
  • Clearer error messages: parameter validation failures return earlier, and errors for unknown models and exhausted key quotas are more explicit.

New models this month (nearly 30)

Chat / general

  • claude-opus-5: Anthropic flagship model with a 1M context window and 128K max output, built for heavy reasoning, coding and long-horizon agent tasks.
  • claude-sonnet-5: for chat, reasoning and agent scenarios.
  • gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna: three tiers of the OpenAI GPT-5.6 series, each with a 1,050,000 context window, 128K max output, and text plus image input. Sol is the flagship tier, Terra matches GPT-5.5 performance at half the price, and Luna targets cost-sensitive workloads.
  • qwen3.8-max-preview: the latest generation Qwen foundation model with 2.4T parameters, available at a limited-time 90% discount.
  • kimi-k3: Moonshot AI open long-context model with 2.8T parameters, a 1M context window and native vision input. See the Kimi K3 hands-on guide.
  • grok-4.5: configurable reasoning, tool calling and 500K context, suited to code repair and agent workflows.
  • tencent-hy3: the GA release of Tencent Hunyuan Hy3. MoE architecture with 295B total / 21B active parameters, a 256K context window, open sourced under Apache 2.0.
  • gemini-3.6-flash, gemini-3.5-flash-lite, gemini-2.5-flash-lite: new tiers in the Google Flash series.
  • glm-5.2-fast-preview, Qwen3-235B-A22B-Instruct-2507, command-a-plus-05-2026, gemma-4-31b, longcat-2.0.

Coding

  • qwen3-coder-480b-a35b-instruct: Qwen3-Coder flagship code model for code generation, software engineering agents and tool calling workflows.

Audio

  • gpt-audio-1.5: the first generally available (GA) OpenAI audio model, supporting audio input and output.
  • gpt-4o-transcribe-diarize: ASR transcription with speaker diarization, available only through the Transcription API.
  • qwen-audio-3.0-tts-plus, qwen-audio-3.0-tts-flash: Qwen speech synthesis models. The plus tier offers fine-grained control over emotion and character, the flash tier delivers first-packet latency under 200 ms.

Image

  • mai-image-2.5-pro: Microsoft flagship image generation and editing model with high-fidelity realism and in-image text rendering. mai-image-2.5 and mai-image-2.5-flash from the same series are also live.
  • gemini-3.1-flash-lite-image: Google image capability model.

3D

  • hy-3d-3.1: Tencent Hunyuan 3D professional edition, supporting text-to-3D, image-to-3D and eight-view multi-angle input.

Retrieval / reranking

  • jina-reranker-v3.5: Jina AI multilingual listwise document reranker for RAG, search and structured data ranking.

Pricing and announcements

  • 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.
  • Qwen3.8-Max-Preview at 90% off: calls are billed at 10% of the list price, effectively 10x more usage for the same spend. Limited time, first come first served.
  • DeepSeek V4 time-of-day billing: DeepSeek V4 models support peak and off-peak billing. Charges follow the time window of the request submission time, and streaming requests that cross a boundary are also priced at submission time.

FAQ

Which models were added this month?
Nearly 30, covering chat (claude-opus-5, claude-sonnet-5, the GPT-5.6 series, qwen3.8-max-preview, kimi-k3, grok-4.5, tencent-hy3 and more), coding (qwen3-coder-480b-a35b-instruct), audio (gpt-audio-1.5 and the qwen-audio-3.0-tts series), image (the mai-image-2.5 series), 3D (hy-3d-3.1) and reranking (jina-reranker-v3.5).

How do I connect to the AIHubMix MCP Server?
Add the entry point https://mcp.aihubmix.com/mcp in any MCP-enabled client (Claude Code, Codex, Cursor and others). Credentials are passed per request by the client. You can then query models and prices, search documentation, check your balance, and call chat, image generation and video generation tools.

What do the media generation APIs support?
Submit image generation and editing requests through /ai/v1/images/generations and /ai/v1/images/edits, then use the unified async task API to query status, download artifacts and receive webhook callbacks. See Async Tasks.

What is the Qwen3.8-Max-Preview limited-time offer?
Calls are billed at 10% of the list price, effectively 10x more usage for the same spend. Limited time, first come first served.

Browse all models on the model gallery and find integration details in the documentation.

Updated: 2026-07-31

More from the blog