Unify every model one gateway.

Anthropic's Claude Sonnet 5.5 is the successor to Claude Sonnet 5 — the best combination of speed and intelligence in the lineup. It is the direct upgrade for well-scoped everyday work (building features, fixing bugs, drafting and review) at the same $2 / $10 per million input / output rates, with adaptive thinking on by default at `high` effort, 1M-token context, 128K max output, vision input, and tool use. Faster than Opus 5.5, cheaper, same window.

Step 5 Preview is StepFun's next-generation flagship base model designed for real-world tasks, with a focus on programming and professional expertise. A multimodal (text, image, video in / text out) base model with a 1M-token context and adjustable reasoning effort.

Gemini Embedding 2 is Google's first multimodal embedding model, mapping text, images, video, audio, and PDFs into one 3,072-dimension vector space for cross-modal semantic search, document retrieval, and recommendations over 100+ languages. Upstream accepts 8,192 input tokens and MRL-truncates to 128–3,072 dimensions. AnyRouter exposes the text path only — the OpenAI-compatible /embeddings body is a text `input`, so the extra upstream modalities are deliberately not declared here.

Qwen3 Embedding 4B is the 4B size in the Qwen3 embedding and reranking family, between the 0.6B and 8B cuts. It embeds text over a 32,768-token window at 2,560 dimensions, supports 100+ languages, is instruction-aware for task prefixes, and supports Matryoshka truncation from 32 to 2,560 dimensions. It ranks below the 8B cut on MTEB multilingual but well above the 0.6B cut, at a much lower cost.

Span-01 Lite is a behavior-scoring classifier from Respan — the free, lighter tier of Span-01, a 4B model trained with RLAIF on classification reasoning rather than fixed taxonomies. Send a conversation span and a set of plain-language behavior definitions and get a present / absent / not_observable probability per behavior in a single forward pass, with no token-by-token generation. Respan reports it ahead of Jev 1.13 and Sonnet 5 on their English and multilingual behavior-detection suites. Call POST /api/v1/decisions (POST /api/v1/systemone is the same handler).

Fugu Max is the largest tier of Sakana AI's Fugu line — a 1M-token-context model that takes text, images, and files and returns text. It joins sakana/fugu and sakana/fugu-ultra as a distinct tier, and is not a v2 of fugu-ultra (that one is folded into the existing row as an alias, not a new listing).

Hy-MT2 30B A3B is Tencent's translation model — a 30B-parameter MoE with 3B active parameters, tuned for machine translation with a short 8,192-token window. It is the first translation family in the AnyRouter catalog and the largest of the three published sizes (1.8B, 7B, 30B-A3B); only the 30B-A3B cut is listed here.

Solar Mini 4 is Upstage's small, fast tier in the Solar 4 generation: a text-only model with a 524,288-token context window at roughly $0.05 / $0.20 per million input / output tokens. It is the cheap sibling of upstage/solar-pro4, not a packaging of it, so it gets its own listing row.

GLM 5.3 Prime is the top reasoning tier of Z.ai's GLM 5.3 line, above the Flash and FlashX cuts, with a 1M-token context window. It sits alongside z-ai/glm-5.3 and z-ai/glm-5.3-flash as a separately priced official variant rather than a packaging of them, so it keeps its own listing row.

BAAI's BGE-M3 is a multilingual text embedding model that handles dense, sparse, and multi-vector retrieval in one model, across 100+ languages and an 8,192-token input window. It returns a 1,024-dimension dense vector.

BAAI's bge-multilingual-gemma2 is a compact multilingual text embedding model built on Gemma 2, trained across 100+ languages with an 8,192-token input window. It returns a 3,584-dimension dense vector.

Mistral Small 3.2 is a 24B instruction model with a 131k context window, improved instruction following and function calling over Mistral Small 3.1, and the same Apache 2.0 license.

MiMo V2.6 Flash is Xiaomi's full-modality reasoning model for high-frequency invocations and large-scale professional workloads. It keeps the 1M-token context window and up to 128K output tokens of MiMo V2.6 Pro at roughly a third of the input price, making it the best-balanced pick when reasoning quality matters more than peak capability.

MiMo V2.6 Pro is Xiaomi's trillion-parameter, natively omni-modal flagship reasoning model for complex projects, long-horizon agentic tasks, coding, research, and professional knowledge work. It supports a 1M-token context window with up to 128K output tokens, and keeps the V2.5 API price while improving reasoning and long-task completion.

Kev 4B is a small open-weight decision model from Jared Palmer — a LoRA adapter and pointer head on Qwen3.5-4B-Base. Send state and typed questions (noul / choice / score) and get a calibrated probability per question in one forward pass, not generated text. Call POST /api/v1/decisions (POST /api/v1/systemone is the same handler).

Fastino GLiNER2.5-Decide is a decision model: send application state and typed questions (noul / choice / score) and get structured answers with confidence — not generated chat text. Call POST /api/v1/decisions (POST /api/v1/systemone is the same handler). Upstream, GLiNER2.5 runs schema-driven classifications on an OpenAI-compatible chat-completions API; the gateway adapts the System One envelope for this hop.

LongCat-2.5-Preview is Meituan's sparse mixture-of-experts frontier model — ~1.6T total parameters with ~48B active — over a 1M-token context window. It adds native multimodal understanding (image input), stronger coding ability, and a `thinking` toggle, and is built for long-horizon agentic work across terminals, browsers, GUIs, spreadsheets, and design tools. Served over the OpenAI-compatible and Anthropic-compatible APIs.

Inception Labs' Mercury 2.5 is a diffusion language model for fast text generation and reasoning, with a 260,000-token context window.

Llama Guard 4 is a 12B natively multimodal safety classifier for content safety classification on LLM inputs and responses. It acts as an LLM itself: it generates text indicating whether a prompt or response is safe or unsafe and, if unsafe, lists the violated categories (MLCommons hazards taxonomy S1–S14). Use it as a guardrail/judge hop in front of or behind chat models.

OpenAI's GPT-6 Luna is the efficient GPT-6 tier for focused, high-throughput tasks, with configurable reasoning, tool use, image input, and a 1.05M-token context window.

OpenAI's GPT-6 Sol is built for complex coding and agentic workflows, with configurable reasoning, tool use, image input, and a 1.05M-token context window. It is the high-end GPT-6 tier below GPT-6 Astra.

OpenAI's open-weight safety reasoning model built on gpt-oss-20b (Apache 2.0). A 21B-parameter MoE tuned for safety tasks: content classification, LLM input/output filtering, and trust & safety judgments with reasoning traces. Use it as a guardrail/judge hop in front of or behind chat models.

Qwen3-235B-A22B-Instruct-2507 is Qwen's updated non-thinking MoE model for multilingual instruction following, coding, tool use, and long-context work. It activates 22B of 235B parameters and provides a native 262K context window.

L3.1 Euryale 70B is Sao10k's text-generation model for creative roleplay, with a 131K context window. It is the successor to Euryale L3 70B v2.1.