Browse models

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, the Anthropic-compatible /v1/messages route, and the Bedrock wire-format routes. Pick a modality, narrow by maker or capability, and every card gives you the id, the aliases that resolve to it, and a request you can paste straight into a terminal.

Prefer a plain list? The same catalog is published as static pages per modality — Text & chat, Image, Video, Audio & voice, Embeddings — and every model has its own page under the Models tab with SDK examples, limits and same-weights alternatives.

The live catalog

The browser above is generated from the same registries the gateway routes from and is refreshed with every model change. At runtime, treat the API as the source of truth:

curl https://api.aireserve.com/v1/models \
  -H "Authorization: Bearer $AIRESERVE_API_KEY"

GET /v1/models returns the OpenAI-shaped list, filtered to what your organization can call (China-hosted and consent-gated models appear only once an administrator has opted in), so model pickers and framework integrations that enumerate models work unchanged. Retired model IDs return 404 ResourceNotFoundException — enumerate /v1/models rather than hardcoding lists.

How to read a model id

You see	It means
`claude-opus-5-5`, `gpt-5.5`, `wan-2.6`	The canonical gateway id — the string to send as `model`. One id works on every route.
`claude-sonnet`, `claude-opus`	Recommended-release aliases that track our current Sonnet and Opus picks, so pinned integrations opt into upgrades by using the alias instead of a dated id.
`sonnet`, `gpt`, `maverick`, `grok-4.5`, `us.anthropic.…`	Aliases — tier words, family defaults, nicknames, xAI’s dotted ids, AWS Bedrock ids and retired ids all resolve at the gateway; each model page lists the ones that land on it.
`glm-5.3-deepinfra`, `claude-opus-5-5-bedrock`	A serving-provider suffix pins the same weights to one disclosed provider (details).
`deepseek-v4-flash:floor`, `…:nitro`	A routing variant lets the gateway pick the cheapest or fastest provider per request (details).
`claude-sonnet-4-6[1m]`	The 1M-context form of an extended-context model (details).
`claude-air-gpt-5.5`	The Claude Code picker id for a non-Claude model — Claude Code only lists ids that start with `claude`, so the gateway advertises every other model under this self-disclosing prefix.

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.

Family-level notes — the per-model ids, prices and limits live in the browser above and on Text & chat models.

Family	Notes
OpenAI GPT	Vision on 4.1/4o/5.x/6; automatic prompt caching (not on the `-pro` tiers). The GPT-6 line has 1M context plus image and file input. The `-ultrafast` IDs serve the same GPT-6.1 Sol / GPT-6 Astra weights on OpenAI’s fastest service tier (up to 8x faster token generation) at 6x the standard price on every token class — pick them when speed matters more than cost.
Anthropic Claude	Vision; explicit prompt caching (minimums) — `claude-fable-5-1` cache reads cost 2.5% of input (75% below Fable 5); `claude-opus-5-5` (1M context) is the current Opus flagship, served first-party by Anthropic; `-bedrock` and `-vertex` serving variants available (below).
Google Gemini	Vision + native video input (table); implicit caching.
xAI Grok	Automatic caching; `grok-4-20-non-reasoning` serves the 4.20 weights without extended reasoning.
DeepSeek	`deepseek-v4.1-flash` is DeepSeek V4.1 Flash on US-hosted DeepInfra; `deepseek-flash` serves the same new weights on DeepSeek’s official API and requires the China-hosting and training consents. V4.1 Flash is the only released V4.1 model — there is no separate non-Flash V4.1 endpoint. `deepseek-v3.2` is US-served by DeepInfra (DeepSeek retired it from its own API) with a `deepseek-v3.2-bedrock` AWS route and a consent-gated `-streamlake` variant; `deepseek-v4-flash-vision-exp-novita` is an experimental vision build served by Novita AI with same-weights `-deepinfra` serving variants.
Meta Llama	Served on OpenRouter and AWS Bedrock (same weights, load-balanced); `-bedrock` variants pin the Bedrock route.
Mistral	`mistral-large` served on OpenRouter and AWS Bedrock (same weights, load-balanced); `mistral-large-bedrock` pins the Bedrock route.
Moonshot / Kimi	`kimi-k3` and `moonshot-v1-*` are China-hosted (below) — `kimi-k3-fireworks` is the US-served variant. The coding-tuned K2.7 Code serves via `kimi-k2.7-code-moonshot` (China-hosted, consent-gated) and the `-deepinfra` / `-novita` variants; the old US-served `kimi-k2.7-code` ID still resolves (below).
Alibaba Qwen	`qwen-3.8-max` and `qwen-3.8-flash` are Alibaba-served (China-hosted, below); `qwen-3.8-max-deepinfra` / `-novita` are the US-served variants.
Perplexity Sonar	Web-grounded answers with citations.
NVIDIA Nemotron	Open-weight, served by DeepInfra (US); the opt-in `nemotron-3-ultra-nvidia` variant sits behind an organization training consent.
Other open-weight	`minimax-m3`, GLM 5.x, GLM 4.7 and `gpt-oss-*` on managed US hosting; serving variants below. `glm-5.3-flash-novita` is Z.AI’s GLM-5.3-Flash (320B-A18B, 1M context, image input) served by Novita AI; `glm-4.7-novita` is Z.AI’s GLM 4.7 flagship (200K context, text only) with same-weights `glm-4.7-deepinfra` and `glm-4.7-bedrock` variants.

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; and when Baseten deprecated its GLM 4.7 deployment in October 2026, glm-4.7-baseten was repointed at the same weights on Novita AI (glm-4.7-novita, same price and context window). 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 and nemotron-3-super-nvidia since October 2026, when NVIDIA retired the hosted endpoints — the DeepInfra-served nemotron-3-nano and nemotron-3-super remain available; and minimax-m2.7 since September 2026, when Together withdrew its only serving).

Bedrock-routed variants & AWS model IDs

Eighteen models are also offered via AWS Bedrock serving under -bedrock IDs — same shapes, same key, one-string A/B against the direct path — including the current Claude line (claude-fable-5-1-bedrock, claude-fable-5-bedrock, claude-opus-5-5-bedrock, claude-opus-5-bedrock, claude-sonnet-5-5-bedrock, claude-sonnet-5-bedrock, claude-opus-4-8-bedrock, claude-opus-4-7-bedrock). 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. More Vertex AI routes are being verified: the Gemini line, the Claude 4.6+ / 5.x line and Model Garden partner models (xAI Grok, Z.ai GLM, DeepSeek, Meta Llama 4, OpenAI open weights) under the same -vertex suffix. A route appears in this catalog the day its gateway verification passes; until then its id is not listed by GET /v1/models.

Azure-routed variants

Microsoft Azure AI Foundry is being onboarded as a serving provider under -azure IDs: Azure OpenAI deployments of the GPT line, and Foundry Models sold directly by Azure for Anthropic Claude, xAI Grok, DeepSeek, Moonshot Kimi, Meta Llama, Mistral and OpenAI open weights. Microsoft operates the model inside Azure in the United States as a data processor; the model’s publisher does not receive prompts or completions, and Microsoft’s abuse-monitoring retention (up to 30 days on our account type) applies — see the provider data-handling disclosures. No -azure route is listed in this catalog yet: each one appears the day its gateway verification passes.

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, deepseek-v4.1-flash-novita for the same weights served by Novita AI (US), 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-bedrock (all US; the AWS route is the same weights on Amazon Bedrock, see Bedrock-routed variants; the glm-4.7-baseten variant retired with Baseten’s deployment in October 2026 and now resolves to glm-4.7-novita). 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. Same weights, different machines: output speed, time to first token, and prompt-cache behavior are the provider’s, so two suffixes of one model can differ by 2–3× in tokens per second (measured across the gpt-oss-120b legs, October 2026). The unsuffixed ID is the default we picked for speed and reliability; reach for a suffix when you need a specific provider’s data-handling terms or want to compare. The portal’s Analytics Usage by Model table breaks a model out by serving provider (once more than one has served it) with p50 / p95 latency and cache-hit rate on your own traffic, which is the fairest comparison. Since September 2026, GMI Cloud (US) serves seven -gmicloud variants: 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 (deepseek-v3.2-gmicloud retired with GMI’s V3.2 serving in September 2026 and resolves to deepseek-v3.2). GMI Cloud publishes no API-content retention terms yet, so the -gmicloud variants require the organization data-retention consent (see “Organization consent gates” below). Since October 2026, Sail Research serves the -sail variants kimi-k3-sail, glm-5.3-sail, glm-5.3-flash-sail, deepseek-v4.1-flash-1017-sail, deepseek-v4-pro-sail, deepseek-v4-flash-sail, and gemma-4-31b-sail — same weights as each base model — plus two Sail-only entries, gemma-4-12b and gemma-4-31b-nvfp4 (an NVIDIA NVFP4 build of Gemma 4 31B). Sail’s published terms are zero data retention by default (inference prompts and outputs are processed only in memory and never persistently stored) with no training on inference data without written consent, so no organization consent gate applies to these variants; Sail’s processing region is not fixed (multi-region by default) — see the provider data-handling disclosures for its row. Novita AI retires its DeepSeek V4 Flash serving on October 9, 2026 at 16:00 UTC (Novita changelog). deepseek-v4-flash-novita, deepseek-v4-flash-reasoner-novita, deepseek-v4-flash-vision-exp-novita, and gpt-oss-20b-novita keep working until then and are no longer fallback destinations for any other route. Novita’s designated successor is DeepSeek V4.1 Flash, available now as deepseek-v4.1-flash-novita — a different model, so it is never substituted silently; switch to it (or to another serving variant of the same V4 Flash weights, such as deepseek-v4-flash-deepinfra) before the cutoff. Organizations that used the retiring variants recently receive the standard deprecation notice (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:

:auto — your organization’s provider preference from the portal’s Router page: a balance of price, latency and reliability scored per request on live signals (the current rate card, trailing latency, success and rate-limit rates, provider account health), or fixed traffic shares per provider (for example DeepInfra 60% / Baseten 40%), or a pinned provider, plus any excluded providers. The leg that serves a conversation stays sticky so provider prompt caches stay warm; a provider that is down or rate-limited hands its share to the others. Unset, the preference is Balanced.
: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.
:online — not a routing variant but it rides the same place: the gateway runs one live web search on your latest message and hands the model the sources before answering. Combines with one variant (deepseek-v4-flash:online:nitro).

curl https://api.aireserve.com/v1/chat/completions \
  -H "Authorization: Bearer $AIRESERVE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash:auto",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

The response body and the x-aireserve-served-model header name the provider-suffixed ID that actually served (for example deepseek-v4-flash-deepinfra), and x-aireserve-served-provider names the provider, so the serve is always disclosed. Send x-aireserve-route-reason: 1 to get the decision back as a compact response header (preset=balanced; served=deepinfra; scores=deepinfra:0.91,baseten:0.78; …). 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 still applies to whichever provider the variant picks, minus any provider your preference excludes (fallbacks are opt-in per organization or per request; an organization that has at least one fallback chain configured also keeps the platform’s standard provider failover for its other models; ask support to change either). 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.

OpenRouter `provider` block

Migrating from OpenRouter? :nitro and :floor work as-is, and the request-level provider object is honored on /v1/chat/completions — it is consumed by the gateway and never forwarded upstream:

Field	Effect
`sort: "price"` / `"latency"` / `"throughput"`	Cheapest / fastest for this request (overrides the `:auto` preference).
`order: ["Baseten", "DeepInfra"]`	Serve from the first listed provider that can take the request; unlisted providers are scored as usual if none can. Naming a provider is your choice: a provider that serves the model but hasn’t yet earned a place in our automatic routing still serves when you list it.
`allow_fallbacks: false`	Restrict the serve to `order` and send no failover chain, exactly like `fallbacks: []`. If no listed provider can serve, the request is refused with `503 SERVICE_UNAVAILABLE` instead of being rerouted.
`only: ["Baseten", "DeepInfra"]`	Allow-list: serve only from these providers (ranked as usual among them), and fail over only among them. Nothing allowed can serve → `503`.
`ignore: ["Novita"]`	Exclude providers for this request (added to your organization’s exclusions).
`max_price: { "prompt": 0.3, "completion": 1.2 }`	Ceilings in USD per 1M tokens on the provider’s list rate. A provider over either cap is skipped; none within the caps → `503`, and the message quotes the lowest listed rates so you can adjust. `request` and `image` caps are accepted and ignored.
`zdr: true` or `data_collection: "deny"`	Serve only from zero-data-retention providers (today: Together AI, Fireworks AI, DeepInfra, Novita AI, Baseten, Cloudflare Workers AI). Works on every model: a model only a retaining provider serves is refused with `503` rather than served quietly.
`healing: false`	Turn off JSON response healing for this request. Not a routing control: it is honored on its own, with or without the rest of the block.
anything else (`require_parameters`, `quantizations`, …)	Accepted and ignored; provider names we don’t serve are dropped silently.

{
  "model": "deepseek-v4-flash",
  "messages": [{ "role": "user", "content": "Hello" }],
  "provider": { "order": ["Baseten"], "allow_fallbacks": false }
}

Restrictions (only, allow_fallbacks: false, max_price, zdr / data_collection) are contracts: the gateway honors them or refuses the request, never widens around them — and a provider they exclude is also removed from the request’s failover chain. A 503 from a restriction carries the restriction’s name in the error message. Preferences (sort, ignore, a bare order) never fail a request. A provider block that sets no sort or order keeps the suffix’s meaning, so deepseek-v4-flash:floor with { "provider": { "zdr": true } } is the cheapest zero-retention provider. Provider names are matched case-insensitively against our serving-provider labels (baseten, deepinfra, together, novita, fireworks, openai, anthropic, bedrock, vertex, gemini, grok, mistral, …); OpenRouter display names such as "Together AI", "Amazon Bedrock" or "xAI" are accepted too.

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, and the Prodia-served image models flux-2-dev, flux-2-klein-4b, flux-2-klein-9b, flux-1.1-pro and sdxl). 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 one of the three pair routers — NVIDIA NeMo Switchyard, NVIDIA Switchyard Escalation, or RouteLLM — 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; every router listed is live. The separate Models tab controls model availability and routing eligibility. When enabled, the gateway may serve a turn on the efficient model the administrator chose instead of the model the request named. The three routers differ in how they decide:

NVIDIA NeMo Switchyard (stage routing) serves a routine turn of a coding-agent session (Claude Code, Codex, and other tool-calling harnesses) on the efficient model and brings 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. 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.
NVIDIA Switchyard Escalation starts every coding-agent session on the efficient model, from its first turn. On each later turn a trajectory judge — AI Reserve’s internal routing model, reading a condensed view of the task framing and the most recent turns, never the full conversation — rules whether the run is in sustained trouble (loops, false progress, drift, giving up); the number of consecutive trouble verdicts the administrator sets (upstream’s benchmarked default is two) moves the rest of the session to the requested model, and a session that has escalated stays there with no further judge calls. It is a TypeScript port of Switchyard’s open-source escalation router, with one stated deviation: the judge reads the trajectory as it stands when a request arrives rather than after a buffered reply from the efficient model, so nothing is served twice. The judge call is routing overhead AI Reserve pays for; it adds a short delay to each unlatched turn.
RouteLLM scores every chat request — pinned models included — with the open-source RouteLLM matrix-factorization router from LMSYS, run in-process with its published checkpoint: an embedding of the latest prompt predicts how likely the requested model is to answer better than the efficient one, and prompts below the cost threshold go to the efficient model. The embedding (OpenAI text-embedding-3-small, the model the checkpoint was trained against) is routing overhead AI Reserve pays for, and it means each routed prompt reaches OpenAI: an organization that blocks OpenAI as a maker or whose processing-region policy excludes it is not routed by RouteLLM at all — every request gets the model it asked for, and the Routing tab says so. The standard interface calibrates the threshold on the organization’s own traffic to the strong-model share its dial asks for, falling back to RouteLLM’s published Chatbot Arena calibration until enough requests have been seen; the native knob pins the threshold itself.

First-party conversational surfaces (Microsoft Office, the portal chat, and MCP) are excluded from all three outright, whatever their turns carry. The Slack agent is excluded from the escalation and RouteLLM routers outright as well; under the stage router its turns are in scope only when they carry tool calls, like any coding harness. Each router is an in-process open-source integration, not an NVIDIA- or LMSYS-hosted routing service or a claim of their endorsement. Every switched response is disclosed: the x-aireserve-served-model header names the model that served, and the usage record carries the requested model, the served model, and which router decided. 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 from organizations that have not opted in are unaffected; requests with model: "auto" first receive their frontier model from the organization’s pool, which then plays the requested model’s part.

Benchmark Router (organization opt-in)

An enterprise administrator can route every request that asks for model: "auto" to the enrolled model with the top published benchmark score, by choosing Benchmark Router as the organization’s model router on the Routing tab. The ranking is deterministic arithmetic over independent benchmark results published by Artificial Analysis and refreshed daily — no model is called to decide. In standard tuning the pool is ranked on the Artificial Analysis Intelligence Index, discounted by the control plane’s cost weight (a price term across the pool) and speed weight (a latency term); in native tuning the administrator weights up to three benchmarks (Intelligence, Coding, Math and Agentic indices, Humanity’s Last Exam, MMLU-Pro, GPQA Diamond, GDPval, LiveCodeBench, IFBench, τ²-bench, SWE-bench) summing to 100 and nothing else applies. A model with no published score for a weighted benchmark is left out of the ranking; the Routing tab shows, per enrolled model, which scores exist, and when nothing can be ranked AI Reserve picks from the routed pool instead. The block is listed only in environments where benchmark data has been loaded. Requests that name a model are served as asked.

Not Diamond as your model router (organization opt-in)

An enterprise administrator can route every request that asks for model: "auto" through Not Diamond’s commercial model router under the organization’s own Not Diamond account, by choosing Not Diamond as the organization’s model router on the Routing tab and pressing Test & save with the organization’s Not Diamond API key. The key is kept only once Not Diamond answers a real test call with a model the gateway serves; it is stored encrypted, never shown again, and Not Diamond usage is billed to the organization’s own Not Diamond account. Each routing call sends Not Diamond the organization’s enrolled models (as Not Diamond names them, or described by price and context length for models it does not list) and the latest user message — hashed by default, so Not Diamond routes on a hash and does not persist the text; never the system prompt, earlier turns, tool results, or user identities. Not Diamond’s answer is checked against the organization’s model policy before it serves, and when Not Diamond does not answer in time AI Reserve picks from the routed pool instead. Requests that name a model are never sent to it. Tuning follows either the control plane’s dial (cost ↔ quality becomes Not Diamond’s cost_quality_tradeoff, a speed-leading dial its latency tradeoff) or Not Diamond’s own knobs, including a custom router trained on the organization’s evaluations (preference_id). This is a bring-your-own-account integration, not a Not Diamond service operated by AI Reserve or a Not Diamond endorsement.

Your own routing endpoint (organization opt-in)

An enterprise administrator can make a service you run pick the model for every request that asks for model: "auto", by choosing Your endpoint as the organization’s model router on the Routing tab, entering the endpoint’s URL and HTTP method, and pressing Test & save. The endpoint is saved only once it answers the test with a model the gateway serves; until then (and whenever it does not answer in time) AI Reserve picks from the organization’s routed pool. Requests that name a model are never sent to it. Request. Every routing call sends the organization’s routed models as gateway model IDs, the administrator’s own routing preferences (omitted when empty), and the conversation id when the request carries one:

POST https://router.example.com/route
Content-Type: application/json

{
  "models": ["claude-opus-4-8", "gpt-5.5", "gemini-3.5-flash"],
  "preferences": { "region": "us" },
  "conversationId": "conv_8f2…"
}

GET https://router.example.com/route?models=claude-opus-4-8,gpt-5.5&preferences=%7B%22region%22%3A%22us%22%7D&conversationId=conv_8f2…

Tune with. The block offers the two tuning modes every router on the Routing tab has:

Native (the default, and how every endpoint saved before this option existed behaves): the request above, with only the preferences JSON the administrator entered.
AI Reserve standard: the organization’s control plane — the quality / cost / speed dial and the plain-language routing instructions set under the AI Reserve Prompt Router block — is added to preferences as aiReserve on every routing call, beside the administrator’s own keys:

{
  "models": ["claude-opus-4-8", "gpt-5.5", "gemini-3.5-flash"],
  "preferences": {
    "region": "us",
    "aiReserve": {
      "dial": { "smarter": 45, "cheaper": 30, "faster": 25 },
      "instructions": "Prefer models with strong tool use. Keep routine coding requests cost efficient."
    }
  }
}

dial carries the three stored percentages (smarter is the quality share, cheaper the cost share, faster the speed share; they sum to 100) and is null until the organization has saved a dial; instructions is the instructions text or null when none are set. Both keys are always present in standard mode, and aiReserve is reserved — a same-named key in the administrator’s own preferences is replaced. On a GET endpoint the same object arrives URL-encoded in the preferences query parameter. The Routing tab shows the exact request either mode sends, and Test & save sends that same request, so an endpoint is never saved against a shape it has not answered. Response. Answer within the deadline (2.5 seconds per routing call; the administrator’s test waits up to 5 seconds) with a model identifier as JSON — {"model": "gpt-5.5"}, with modelId or id also accepted — a bare JSON string, or a single-line plain-text body. Loosely formatted names are tolerated (Opus-4-8 resolves), but the model must be served on the gateway and permitted by the organization’s model policy and the requesting key’s quota; any other answer, or no answer in time, falls back to the organization’s routed pool for that request. The x-aireserve-served-model header on the response names the model that served.

Image generation

Every image model — generation via POST /v1/images/generations and editing via POST /v1/images/edits — is listed with pricing on Image models and in the browser above under Image. The image generation guide covers parameters, response shapes and the async edit job flow. The Qwen Image 3 models and seedream-5 (ByteDance’s official Ark API) are China-hosted (see above) and follow the same per-client opt-in as the chat models. 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

Every video model on the async POST /v1/videos/generations job API — text-to-video, image-to-video and reference-to-video — is listed with per-second pricing on Video models and in the browser above under Video. The video generation guide covers inputs, durations, resolutions and polling. 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

Capability	Where documented
Streaming	Streaming — all chat models
Tool use / function calling	Tool use
Vision (image input)	Images — per-model table
Video input	Video input — per-model table
Prompt caching	Prompt caching — per-model minimums
Token counting	`POST /v1/messages/count_tokens` (Anthropic-shape)

Overview

By maker

The live catalog

How to read a model id

Chat & reasoning models

Deprecation policy

Bedrock-routed variants & AWS model IDs

Vertex-routed variants

Azure-routed variants

Provider-suffixed serving variants

Provider routing variants

OpenRouter `provider` block

1M-token context

China-hosted models

Tier switching (organization opt-in)

Benchmark Router (organization opt-in)

Not Diamond as your model router (organization opt-in)

Your own routing endpoint (organization opt-in)

Image generation

Video generation

Capability reference

​The live catalog

​How to read a model id

​Chat & reasoning models

​Deprecation policy

​Bedrock-routed variants & AWS model IDs

​Vertex-routed variants

​Azure-routed variants

​Provider-suffixed serving variants

​Provider routing variants

​OpenRouter provider block

​1M-token context

​China-hosted models

​Organization consent gates

​Tier switching (organization opt-in)

​Benchmark Router (organization opt-in)

​Not Diamond as your model router (organization opt-in)

​Your own routing endpoint (organization opt-in)

​Image generation

​Video generation

​Capability reference

The live catalog

How to read a model id

Chat & reasoning models

Deprecation policy

Bedrock-routed variants & AWS model IDs

Vertex-routed variants

Azure-routed variants

Provider-suffixed serving variants

Provider routing variants

OpenRouter `provider` block

1M-token context

China-hosted models

Organization consent gates

Tier switching (organization opt-in)

Benchmark Router (organization opt-in)

Not Diamond as your model router (organization opt-in)

Your own routing endpoint (organization opt-in)

Image generation

Video generation

Capability reference