Skip to main content
One API key unlocks every model below through a single, flat ID namespace — the same model / modelId string works across the Converse SDKs, the OpenAI-compatible endpoints, and the Bedrock wire-format routes. The live catalog is always available programmatically:
GET /v1/models returns the OpenAI-shaped list, so model pickers and framework integrations that enumerate models work unchanged. Treat it as the source of truth — the tables below are a human-readable snapshot.

​
Chat & reasoning models

Reasoning models and small max_tokens. Models that think before answering (gemini-3-flash-preview, gpt-5.x, gpt-6-astra, grok-4-1-fast-reasoning, …) spend part of the token budget on internal reasoning. With a tight cap (e.g. max_tokens: 50) the budget can be consumed before any visible text is produced — the response then comes back with finish_reason / stop_reason = max_tokens (or length) and empty content. That’s the model, not an error: raise max_tokens (a few hundred is a safe floor for reasoning models) or pick a non-reasoning model for short outputs.
Convenience aliases: claude-sonnet and claude-opus track our recommended current Sonnet and Opus releases, so pinned integrations can opt into upgrades by using the alias instead of a dated ID. Retired model IDs return 404 ResourceNotFoundException — enumerate /v1/models rather than hardcoding lists.

​
Deprecation policy

No surprise deprecations: you hear it from us first.
  • At least 30 days of notice. When a model must leave the catalog, the retirement date is announced a minimum of 30 days ahead. Until that date the model keeps working normally.
  • How notices arrive. Your organization’s admins receive an email naming the model, the retirement date, and what to use instead. It goes to every organization that used the model in the previous 60 days, and the same notice stays visible on the admin Models tab until the date passes.
  • Upstream retirements are absorbed where possible. When a provider retires a model or an alias on their side, we keep the catalog ID working wherever we can by pointing it at the successor. Precedent: DeepSeek retired its deepseek-chat and deepseek-reasoner upstream aliases in July 2026, and both IDs kept working unchanged on this platform. Each now calls the explicit DeepSeek V4 successor with its original thinking behavior preserved. Likewise, when Together AI withdrew serverless serving of several open-weight deployments in September 2026, the affected IDs were repointed at the same weights on a surviving provider: kimi-k2.7-code → kimi-k2.7-code-moonshot, kimi-k2.6-together → kimi-k2.6-moonshot (both China-hosted and consent-gated — the re-route serves only after your organization’s opt-ins, below), minimax-m2.5 → minimax-m2.5-tensormesh (consent-gated), and nemotron-3-ultra-together → nemotron-3-ultra. A deprecation notice is sent only when an ID truly cannot be preserved; an ID retired upstream with no successor returns an explicit “retired by its provider” error rather than a generic 404 (as mistral-small does since September 2026, when OpenRouter withdrew its last endpoint; nemotron-3-nano-nvidia since August 2026, when NVIDIA retired the hosted endpoint — the DeepInfra-served nemotron-3-nano remains available; and minimax-m2.7 since September 2026, when Together withdrew its only serving).

​
Bedrock-routed variants & AWS model IDs

Twelve models are also offered via AWS Bedrock serving under -bedrock IDs — same shapes, same key, one-string A/B against the direct path. See Bedrock-routed models for the full mapping. On the Bedrock wire-format routes (/model/{id}/converse[-stream]) the gateway additionally accepts native AWS model IDs (us.anthropic.…, inference-profile ARNs) so unmodified boto3 code works without renaming models — see AWS SDK (boto3).

​
Vertex-routed variants

Three Claude models are additionally offered via Google Vertex AI serving under -vertex IDs — claude-sonnet-4-6-vertex, claude-sonnet-4-5-vertex, and claude-haiku-4-5-vertex — same shapes, same key, one-string A/B against the direct path. Google operates the model inside Google Cloud infrastructure in the United States (us-east5) as a data processor; the model’s publisher does not receive prompts or completions. See the provider data-handling disclosures for the full routing notes.

​
Provider-suffixed serving variants

Several open-weight models are served by more than one disclosed provider. DeepSeek V4.1 Flash uses deepseek-v4.1-flash for its US-hosted DeepInfra primary and deepseek-flash for the separately consent-gated DeepSeek official endpoint. A provider suffix pins other requests to a provider’s deployment of the same weights — for example deepseek-v4-flash-deepinfra / deepseek-v4-flash-novita (DeepInfra and Novita AI serving of DeepSeek V4 Flash), qwen-3.8-max-deepinfra / qwen-3.8-max-novita, glm-5.3-flash-baseten, glm-5.2-together, or kimi-k3-fireworks. Baseten (US) serves -baseten variants of GLM 5.3 (glm-5.3-baseten), GLM 5.2 (glm-5.2-baseten), DeepSeek V4 Pro (deepseek-v4-pro-baseten), Kimi K2.6 (kimi-k2.6-baseten), and Kimi K2.7 Code (kimi-k2.7-code-baseten). Z.AI’s GLM 4.7 (200K context) serves entirely through provider-suffixed variants — glm-4.7-novita (primary), glm-4.7-deepinfra, and glm-4.7-baseten (all US). The -streamlake variants (deepseek-v3.2-streamlake, deepseek-v4-flash-streamlake, deepseek-v4-pro-streamlake, glm-5.2-streamlake, kimi-k2.6-streamlake, kimi-k2.7-code-streamlake, minimax-m2.5-streamlake) are served by StreamLake, Kuaishou’s international AI cloud — they are labeled China-based and additionally consent-gated (see China-hosted models and Organization consent gates below). The suffix names the serving provider whose data-handling terms apply — the model’s original publisher never receives requests on these routes. Unsuffixed open-weight IDs default to a disclosed US serving provider — for example the Nemotron 3 family (nemotron-3-super, nemotron-3-nano, nemotron-3-ultra) has been served by DeepInfra since August 2026. Since September 2026, GMI Cloud (US) serves eight -gmicloud variants: deepseek-v3.2-gmicloud, deepseek-v4-flash-gmicloud, deepseek-v4-pro-gmicloud, glm-5.2-gmicloud, glm-5.3-gmicloud, glm-5.3-flash-gmicloud, kimi-k2.6-gmicloud, and kimi-k2.7-code-gmicloud — same weights as each base model, often at the lowest US-hosted rate for those weights. GMI Cloud publishes no API-content retention terms yet, so the -gmicloud variants require the organization data-retention consent (see “Organization consent gates” below).

​
Provider routing variants

Don’t want to pick a provider suffix yourself? Append a routing variant to the model ID and the gateway picks the serving provider per request:
  • :nitro — the fastest provider currently serving those weights (rolling median latency on live traffic). Alias: :fastest.
  • :floor — the cheapest provider by blended token price. Alias: :cheapest.
The response body and the x-aireserve-served-model header name the provider-suffixed ID that actually served (for example deepseek-v4-flash-deepinfra), so the serve is always disclosed. Variants never change what you’re allowed to use: access rules, consent gates, and model allowlists apply exactly as they do without the suffix. Your failover protection also survives the reroute: the model’s fallback chain (your configured chain, or the platform default) still applies to whichever provider the variant picks. On models with a single serving provider, on providers we haven’t yet verified with live production traffic, and on the China-hosted models — which always serve exactly the provider their ID promises — the suffix is accepted and simply serves the model unchanged. Migrating from OpenRouter? :nitro and :floor work as-is.

​
1M-token context

The extended-context models — claude-sonnet-4-6, fable-5, and opus-4-8 — serve a full 1M-token window through the gateway. Two equivalent ways to opt in:
  • Append the [1m] suffix to the model ID (e.g. claude-sonnet-4-6[1m]) — the gateway accepts this form directly.
  • Send the anthropic-beta: context-1m-2025-08-07 header, which the gateway relays to Anthropic.
Tokens are billed at the model’s normal rates. Running Claude Code? It budgets a 200K window by default behind a custom base URL — pin the [1m] form through ANTHROPIC_MODEL to get the full window; see Why the [1m] suffix? for the details.

​
China-hosted models

Some catalog models are served by providers operating under PRC jurisdiction: deepseek-chat and deepseek-reasoner (DeepSeek’s own API), the Moonshot official-API models (kimi-k3, kimi-k2.5, kimi-k2-thinking, moonshot-v1-8k, moonshot-v1-32k, moonshot-v1-128k), the Alibaba Cloud Model Studio-served entries (qwen-3.8-max and the other Model Studio-served Qwen IDs, plus the -alibaba suffixed serving variants), the -deepseek / -moonshot suffixed serving variants, the -streamlake suffixed serving variants (StreamLake is Kuaishou’s international AI cloud — data storage is stated as Singapore, but the operator’s parent is PRC-headquartered and its inference hosting location is unstated), and the ByteDance-served image and video models. The distinguishing test is who operates the serving endpoint, not the model’s origin — Chinese-origin open weights served by US providers (for example deepseek-v4-pro, qwen-3.7-max, kimi-k2.6, or deepseek-reasoner-bedrock on AWS) are not China-hosted. Note that Alibaba’s Qwen chat endpoint runs in the United States (Virginia) for every Qwen chat model except qwen-3-coder-next, which Alibaba serves only from its international (Singapore) endpoint; all of these models are labeled China-based because the operator is PRC-headquartered. Organization admins control access with a per-client opt-in: when disabled, China-hosted models are excluded everywhere — the chat picker, auto-routing, and API traffic through the gateway. See the provider data-handling disclosures for what each provider receives. Separately from the China-hosted opt-in, models served by providers whose terms permit training on API inputs — DeepSeek’s own API, Moonshot’s own API, the NVIDIA-served -nvidia Nemotron variants, and the StreamLake-served -streamlake variants (StreamLake’s international Terms of Service license it to use inputs and outputs for model training, with no published opt-out) — are locked for every organization until an enterprise administrator explicitly opts in from the admin console. The -streamlake variants stack both gates: the China-hosted opt-in and the training consent are each required. The same applies to models whose provider retention cannot yet be verified or disabled (the -tensormesh and -gmicloud serving variants). Locked models appear greyed-out in the model catalog with the required consent named, are excluded from chat and automatic routing, and API requests to them are refused. The provider data-handling disclosures carry the full consent-gate matrix.

​
Tier switching (organization opt-in)

Off for every organization until an enterprise administrator turns it on from the admin console’s Routing tab, by choosing NVIDIA NeMo Switchyard as the organization’s model router and picking the efficient model. The Routing tab is an open marketplace of routers under the AI Reserve control plane: one model router and one provider router per organization, each tunable through AI Reserve’s standard quality / cost / speed dial or its own native settings, with routers that are not yet available shown greyed out. The separate Models tab controls model availability and routing eligibility. When enabled, the gateway may serve a routine turn of a coding-agent session (Claude Code, Codex, and other tool-calling harnesses) on the efficient model the administrator chose instead of the model the request named, and bring the requested model back when the conversation shows trouble — repeated or serious tool errors, a stalled session, or a compacted context. Only requests that already carry tool calls or tool results are candidates: a plain chat message or single-turn API call is never switched, and non-coding surfaces (Microsoft Office, the portal chat, and the Slack agent) are excluded outright, whatever their turns carry. The decision is a deterministic read of the tool calls and results already in the request (a TypeScript port of NVIDIA NeMo Switchyard’s open-source stage router), steadied on a best-effort basis within a coding session so tiers are not flipped back and forth on borderline turns (the steadying memory lives on the gateway instance that served the earlier turn, so it is not guaranteed on every turn); no request content leaves the gateway to make it, and the per-session memory is a tier label — never conversation content. This is an in-process open-source integration, not an NVIDIA-hosted routing service or a claim of NVIDIA endorsement. Every switched response is disclosed: the x-aireserve-served-model header names the model that served, and the usage record carries both the requested and the served model. A switch never crosses your governance: the efficient model must itself be routable for your organization (not blocked, consent-gated, or restricted) and permitted by the requesting key’s quota allow- and blocklists; turns carrying images or files anywhere — including screenshots returned by tools — are never switched; and a turn is left alone when the efficient model could not take it (a thinking setting it does not support, a context larger than its window) or would not be at least 2× cheaper on both input and output. Requests with model: "auto" and requests from organizations that have not opted in are unaffected.

​
Image generation

Via POST /v1/images/generations (guide & pricing): gemini-2.5-flash-image, gemini-3.1-flash-image, gpt-image-1, qwen-image-3.0, qwen-image-3.0-pro, and the fal-served nano-banana-2, flux-2-flex, and grok-imagine-image. The Qwen Image 3 models are China-hosted (see above) and follow the same per-client opt-in as the chat models. Image editing via POST /v1/images/edits (guide): nano-banana-2-edit, flux-2-flex-edit, qwen-image-2-pro-edit, grok-imagine-image-edit, bria-background-remove, and topaz-image-upscale, all served by fal in the United States. The retired dall-e-3, imagen-3, imagen-4, imagen-4-fast, and imagen-4-ultra IDs return errors — migrate to gemini-3.1-flash-image or gpt-image-1.

​
Video generation

Via the async POST /v1/videos/generations job API (guide & pricing): wan-2.6, wan3.0-video, seedance-2.0, seedance-2.0-bytedance, seedance-2.0-mini, seedance-2.0-fast, and seedance-2.5, plus the fal image-to-video / reference-to-video models grok-imagine-video-1.5-i2v, grok-imagine-video-1.5-reference, grok-imagine-video-i2v, flux-3-i2v, seedance-2.5-reference, seedance-2.0-mini-reference, wan-3.0-prime-reference, and minimax-h3-max-i2v. wan3.0-video (Alibaba Model Studio) and the ByteDance official-API models (seedance-2.5, seedance-2.0-bytedance, seedance-2.0-mini, seedance-2.0-fast) are China-hosted (see above); wan-2.6, seedance-2.0, and every fal image-to-video / reference-to-video model are served by fal in the United States.

​
Capability reference