Catalog
Models
One catalog across every provider. Model IDs are stable strings you can hardcode.
The live catalog is the source of truth
GET /v1/models returns the CHAT models in the live catalog (the same list for everyone — it is not filtered by account), including any added since this page was written. Speech and image models are not in that list — see List models.
By modality#
Text
Chat, reasoning, tool calling and vision input.
Speech
Text to speech, billed per character of input.
Image
Generation and editing through an asynchronous job API.
Model families#
Every family is reachable through both request formats.
OpenAI
openai/…General-purpose reasoning and coding models. The default choice when you are not sure what to reach for.
Anthropic
anthropic/…Claude models — strong instruction following, long documents and extended thinking.
Z.AI
z-ai/…GLM models. Strong coding performance with a switchable thinking mode.
MiniMax
minimax/…M-series models for agentic work, reachable through the Anthropic or OpenAI dialect.
DeepSeek
deepseek/…Reasoning-first models with strong mathematics and code generation.
Qwen
qwen/…Multilingual models with vision and video understanding.
google/…Gemini models — fast multimodal generalists across several price points.
xAI
x-ai/…Grok models with adjustable reasoning depth, tuned for agentic coding.
Moonshot
moonshotai/…Kimi models — long-context coding, with a graded reasoning scale on K3.
Mistral
mistralai/…European models with strong code generation — Codestral, Devstral and the Ministral line.
Also in the catalog: NVIDIA Nemotron (nvidia/…), Xiaomi MiMo (xiaomi/…), Tencent Hunyuan (tencent/…), Meta (meta/…) and stealth previews (stealth/…) — smaller families that come and go faster than this page is revised. GET /v1/models always has the current list.
Model IDs#
Every ID is vendor/model. Send the whole thing — the bare model name does not resolve.
{ "model": "openai/gpt-5.6-sol" } // correct
{ "model": "gpt-5.6-sol" } // 404 not_foundWhy the prefix is required
Two vendors can ship similarly named models. Resolving a bare name would mean guessing which one you meant, and guessing wrong bills you for a model you never asked for.
Choosing a model#
- Start mid-tier. Most tasks do not need a flagship, and the latency difference is larger than the quality difference.
- Move up only on evidence.Have a case the smaller model gets wrong before paying several times more per token. "It feels better" is not measurable.
- Classification, extraction and routing — the smallest model that passes your tests is almost always right.
- Check capabilities, not names. A model name does not reliably predict vision, tools or context window. See Capabilities.
Next#
- Capabilities — what vision, tools, reasoning and context window mean for your request.
- Access tiers — which models your account can call.
- Pricing & billing — what each modality is measured in.