Providers

Every provider — built-in or self-hosted — lives under the same providers key. There is no separate list for custom endpoints.

import { createGateway } from '@frogbotai/gateway'

const gw = createGateway({
  providers: {
    openai: {},
    'bedrock': {},
    ollama: { baseURL: 'http://localhost:11434/v1' }
  }
})

openai and bedrock are recognized provider names, so they take that provider’s typed config. ollama isn’t a built-in name, so it’s treated as a generic OpenAI-compatible endpoint — the key becomes its dispatch prefix (ollama/llama3.2).

Built-in providers

Alibaba, Anthropic, Anthropic (AWS), AssemblyAI, Azure, Baseten, Amazon Bedrock, Black Forest Labs, ByteDance, Cerebras, Cohere, Deepgram, DeepInfra, DeepSeek, ElevenLabs, fal, Fireworks, Gladia, Google, Google Vertex, Groq, Hugging Face, Hume, Kling AI, LMNT, Luma, Mistral, Moonshot AI, OpenAI, OpenRouter, Perplexity, Prodia, Replicate, Together AI, TypeSafe AI, Vercel AI Gateway, Voyage, xAI. Each ships its own credential shape (apiKey, region, accessKeyId, …) and its envVars list, used for config fallback and CLI env auto-discovery. Bedrock also supports ambient AWS credentials. See Configuration for the full env var reference. Built-in providers route only models in the catalog. Config also accepts an optional models allowlist of unprefixed catalog IDs to narrow that set. The list is enforced after alias canonicalization across HTTP routes and in-process model resolvers. See Restrict built-in models.

Bedrock inference profiles replace the non-invocable bare IDs for Nova 2 Lite, Llama 3.1 8B, and Llama 3.3 70B. Replace amazon.nova-2-lite-v1:0 with global.amazon.nova-2-lite-v1:0, meta.llama3-1-8b-instruct-v1:0 with us.meta.llama3-1-8b-instruct-v1:0, and meta.llama3-3-70b-instruct-v1:0 with us.meta.llama3-3-70b-instruct-v1:0 in model allowlists and requests.

Built-in provider options

A built-in provider entry takes a config object, an empty object {} to use the defaults, or a pre-built instance. Most providers accept these options:

Option	Type	Default	Description
`models`	`string[]`	Every catalog model for the provider	Allow list of unprefixed model IDs. Each one must be in the catalog, or startup fails. See Restrict built-in models.
`apiKey`	`string`	The provider’s credential variable below	The provider’s API key. For providers marked “Startup check”, startup fails when neither `apiKey` nor that variable is set and the provider accepts no other credential form, or when `apiKey` is an empty string. Not accepted by `replicate`, which uses `apiToken`.
`baseURL`	`string`	The provider’s public API URL	URL prefix for API calls, for example a proxy or regional endpoint. Not accepted by `assemblyai`, `deepgram`, `elevenlabs`, `gladia`, `hume`, or `lmnt`.
`headers`	`Record<string, string>`	—	Extra headers sent with every request to the provider.

Each provider’s environment variables are listed below. The first one is its credential. The CLI builds a provider from these variables when they’re set; see Auto-discover from env.

Key	Provider	Environment variables	Credential check
`alibaba`	Alibaba	`ALIBABA_API_KEY`, `ALIBABA_BASE_URL`	Startup check
`anthropic`	Anthropic	`ANTHROPIC_API_KEY`, `ANTHROPIC_BASE_URL`	Startup check (`apiKey`, or `authToken`)
`anthropic-aws`	Anthropic (AWS)	`ANTHROPIC_AWS_API_KEY`, or `AWS_ACCESS_KEY_ID` + `AWS_SECRET_ACCESS_KEY` + `AWS_REGION` (+ optional `AWS_SESSION_TOKEN`); `ANTHROPIC_AWS_WORKSPACE_ID`	At request time
`assemblyai`	AssemblyAI	`ASSEMBLYAI_API_KEY`	Startup check
`azure`	Azure OpenAI	`AZURE_API_KEY`, plus `AZURE_RESOURCE_NAME` or `AZURE_OPENAI_BASE_URL`; `AZURE_API_VERSION`	At request time
`baseten`	Baseten	`BASETEN_API_KEY`, `BASETEN_BASE_URL`	Startup check
`bedrock`	Amazon Bedrock	`AWS_BEARER_TOKEN_BEDROCK`, or `AWS_ACCESS_KEY_ID` + `AWS_SECRET_ACCESS_KEY` + `AWS_REGION` (+ optional `AWS_SESSION_TOKEN`), or `AWS_PROFILE`, or `AWS_ROLE_ARN` + `AWS_WEB_IDENTITY_TOKEN_FILE`	At request time
`black-forest-labs`	Black Forest Labs	`BFL_API_KEY`, `BFL_BASE_URL`	Startup check
`bytedance`	ByteDance	`ARK_API_KEY`, `ARK_BASE_URL`	Startup check
`cerebras`	Cerebras	`CEREBRAS_API_KEY`, `CEREBRAS_BASE_URL`	Startup check
`cohere`	Cohere	`COHERE_API_KEY`, `COHERE_BASE_URL`	Startup check
`deepgram`	Deepgram	`DEEPGRAM_API_KEY`	Startup check
`deepinfra`	DeepInfra	`DEEPINFRA_API_KEY`, `DEEPINFRA_BASE_URL`	Startup check
`deepseek`	DeepSeek	`DEEPSEEK_API_KEY`, `DEEPSEEK_BASE_URL`	Startup check
`elevenlabs`	ElevenLabs	`ELEVENLABS_API_KEY`	Startup check
`fal`	fal	`FAL_API_KEY` (the CLI also reads `FAL_KEY`), `FAL_BASE_URL`	Startup check
`fireworks`	Fireworks	`FIREWORKS_API_KEY`, `FIREWORKS_BASE_URL`	Startup check
`gladia`	Gladia	`GLADIA_API_KEY`	Startup check
`google`	Google	`GOOGLE_GENERATIVE_AI_API_KEY`, `GOOGLE_BASE_URL`	Startup check
`groq`	Groq	`GROQ_API_KEY`, `GROQ_BASE_URL`	Startup check
`huggingface`	Hugging Face	`HUGGINGFACE_API_KEY`, `HUGGINGFACE_BASE_URL`	Startup check
`hume`	Hume	`HUME_API_KEY`	Startup check
`klingai`	Kling AI	`KLINGAI_API_KEY`, or `KLINGAI_ACCESS_KEY` + `KLINGAI_SECRET_KEY`; `KLINGAI_BASE_URL`	Startup check (`apiKey`, or `accessKey` and `secretKey`)
`lmnt`	LMNT	`LMNT_API_KEY`	Startup check
`luma`	Luma	`LUMA_API_KEY`, `LUMA_BASE_URL`	Startup check
`mistral`	Mistral	`MISTRAL_API_KEY`, `MISTRAL_BASE_URL`	Startup check
`moonshotai`	Moonshot AI	`MOONSHOT_API_KEY`, `MOONSHOT_BASE_URL`	Startup check
`openai`	OpenAI	`OPENAI_API_KEY`, `OPENAI_BASE_URL`, `OPENAI_ORGANIZATION`, `OPENAI_PROJECT`	Startup check
`openrouter`	OpenRouter	`OPENROUTER_API_KEY`, `OPENROUTER_BASE_URL`	Startup check
`perplexity`	Perplexity	`PERPLEXITY_API_KEY`, `PERPLEXITY_BASE_URL`	Startup check
`prodia`	Prodia	`PRODIA_TOKEN`, `PRODIA_BASE_URL`	Startup check
`replicate`	Replicate	`REPLICATE_API_TOKEN`, `REPLICATE_BASE_URL`	Startup check (`apiToken`)
`togetherai`	Together AI	`TOGETHER_API_KEY`, `TOGETHER_BASE_URL`	Startup check
`typesafe-ai`	TypeSafe AI	`TYPESAFE_AI_API_KEY`	Startup check
`vercel`	Vercel AI Gateway	`AI_GATEWAY_API_KEY`, `AI_GATEWAY_BASE_URL`	Startup check
`vertex`	Google Vertex	`GOOGLE_VERTEX_API_KEY`, or `GOOGLE_VERTEX_PROJECT` + `GOOGLE_VERTEX_LOCATION`	At request time
`voyage`	Voyage	`VOYAGE_API_KEY`, `VOYAGE_BASE_URL`	Startup check
`xai`	xAI	`XAI_API_KEY`, `XAI_BASE_URL`	Startup check

Some providers accept more options. Each row is written as <key>.<option> and is set under providers.<key>:

Option	Type	Default	Description
`alibaba.videoBaseURL`	`string`	`https://dashscope-intl.aliyuncs.com`	URL prefix for video generation calls.
`alibaba.embeddingBaseURL`	`string`	`https://dashscope-intl.aliyuncs.com/api/v1`	URL prefix for embedding calls.
`alibaba.includeUsage`	`boolean`	`true`	Asks for token usage at the end of streamed responses.
`anthropic.authToken`	`string`	—	Token sent as `Authorization: Bearer`, instead of `apiKey`. Satisfies the startup check on its own. It isn’t read from an environment variable, and setting both `authToken` and `apiKey` fails on startup; with `ANTHROPIC_API_KEY` set, the CLI discovers an `apiKey`, so unset it to use a token.
`anthropic.name`	`string`	`'anthropic.messages'`	Provider name the AI SDK models report.
`anthropic.generateId`	`() => string`	AI SDK default	Generates IDs for the AI SDK’s response items.
`anthropic-aws.region`	`string`	`AWS_REGION`	AWS region. Required; there’s no fallback region.
`anthropic-aws.workspaceId`	`string`	`ANTHROPIC_AWS_WORKSPACE_ID`	Anthropic workspace ID, sent in the `anthropic-workspace-id` header.
`anthropic-aws.accessKeyId`	`string`	`AWS_ACCESS_KEY_ID`	AWS access key ID for SigV4 signing. Ignored when `apiKey` is set.
`anthropic-aws.secretAccessKey`	`string`	`AWS_SECRET_ACCESS_KEY`	AWS secret access key for SigV4 signing.
`anthropic-aws.sessionToken`	`string`	`AWS_SESSION_TOKEN`	AWS session token for SigV4 signing.
`azure.resourceName`	`string`	—	Azure OpenAI resource name, used to build `https://<resourceName>.openai.azure.com/openai/v1`. Set this or `baseURL`; `baseURL` wins when both are set.
`azure.tokenProvider`	`() => Promise<string>`	—	Returns a Microsoft Entra access token. Called on every request, instead of using `apiKey`.
`azure.apiVersion`	`string`	`'v1'`	Azure API version.
`azure.useDeploymentBasedUrls`	`boolean`	`false`	Uses the older `/deployments/<deployment>` URL format.
`baseten.modelURL`	`string`	—	URL of a custom Baseten model deployment (chat or embeddings), instead of the shared Model APIs.
`baseten.performanceClient`	`BasetenPerformanceClientConstructor`	—	The `PerformanceClient` class from `@basetenlabs/performance-client`, which you install yourself, for batched embedding requests.
`bedrock.region`	`string`	`AWS_REGION`	AWS region. CLI discovery uses `us-east-1` when `AWS_REGION` isn’t set with a bearer token or profile.
`bedrock.accessKeyId`	`string`	`AWS_ACCESS_KEY_ID`	AWS access key ID for SigV4 signing.
`bedrock.secretAccessKey`	`string`	`AWS_SECRET_ACCESS_KEY`	AWS secret access key for SigV4 signing.
`bedrock.sessionToken`	`string`	`AWS_SESSION_TOKEN`	AWS session token for SigV4 signing.
`bedrock.credentialProvider`	`() => PromiseLike<{ accessKeyId: string; secretAccessKey: string; sessionToken?: string }>`	The AWS default credential chain	Returns AWS credentials for each request, instead of the static keys. Without `apiKey`, static keys, or this option, the gateway uses the AWS default chain (profiles, SSO, web identity, instance roles), which needs the optional `@aws-sdk/credential-providers` package.
`black-forest-labs.pollIntervalMillis`	`number`	`500` for images, `2000` for video	Milliseconds between job status checks.
`black-forest-labs.pollTimeoutMillis`	`number`	`60000` for images, `600000` for video	How long to wait for a job before failing.
`cohere.generateId`, `google.generateId`, `huggingface.generateId`, `mistral.generateId`	`() => string`	AI SDK default	Generates IDs for the AI SDK’s response items.
`elevenlabs.webSocket`, `google.webSocket`, `openai.webSocket`, `vercel.webSocket`, `vertex.webSocket`, `xai.webSocket`	`WebSocketConstructor`	The runtime’s `WebSocket`	WebSocket class for streaming connections, for runtimes whose built-in `WebSocket` can’t send headers.
`google.name`	`string`	`'google.generative-ai'`	Provider name the AI SDK models report.
`klingai.accessKey`	`string`	`KLINGAI_ACCESS_KEY`	Kling AI access key. Used with `secretKey` when there’s no `apiKey`.
`klingai.secretKey`	`string`	`KLINGAI_SECRET_KEY`	Kling AI secret key.
`klingai.apiKey`	`string`	`KLINGAI_API_KEY`	Kling AI API key. Satisfies the startup check on its own and takes precedence over `accessKey` and `secretKey`. The CLI discovers it before the access key pair.
`openai.name`	`string`	`'openai'`	Provider name the AI SDK models report.
`openai.organization`	`string`	`OPENAI_ORGANIZATION` in CLI discovery	OpenAI organization ID.
`openai.project`	`string`	`OPENAI_PROJECT` in CLI discovery	OpenAI project ID.
`openrouter.appName`	`string`	`'FrogBot'`	Sent as the `X-OpenRouter-Title` attribution header.
`openrouter.appUrl`	`string`	`'https://www.frogbot.ai'`	Sent as the `HTTP-Referer` attribution header.
`openrouter.compatibility`	`'strict' \| 'compatible'`	`'compatible'`	`strict` sends newer request fields, such as stream options, that only OpenRouter’s own API accepts.
`openrouter.extraBody`	`Record<string, unknown>`	—	Extra fields added to every request body.
`openrouter.api_keys`	`Record<string, string>`	—	Your own API keys for upstream providers, keyed by OpenRouter provider slug such as `anthropic`.
`openrouter.decisionsBaseURL`	`string`	`https://openrouter.ai/api/alpha`, or `baseURL` with `/alpha` in place of a trailing `/v1`	URL for OpenRouter’s Decisions API, used for evaluation models. Set it when `baseURL` is a proxy path that doesn’t end in `/v1`.
`openrouter.baseUrl`	`string`	—	Deprecated spelling of `baseURL`.
`replicate.apiToken`	`string`	`REPLICATE_API_TOKEN`	Replicate API token. Replicate doesn’t accept `apiKey`.
`vercel.headers`	`Record<string, string>`	`http-referer` and `x-title` attribution headers	Extra headers. Your `http-referer` or `x-title` value replaces FrogBot’s, matched without regard to case.
`vercel.teamIdOrSlug`	`string`	—	Vercel team ID or slug, for tokens that can access more than one team.
`vercel.metadataCacheRefreshMillis`	`number`	`300000`	How often Vercel’s model metadata cache refreshes, in milliseconds.
`vertex.project`	`string`	`GOOGLE_VERTEX_PROJECT`	Google Cloud project.
`vertex.location`	`string`	`GOOGLE_VERTEX_LOCATION`	Google Cloud location, such as `us-central1`.
`vertex.anthropic`	`{ location?: string }`	`vertex.location`	Settings for Claude models, which share the project and credentials. `location` overrides the shared location, since some Claude models aren’t served from `global`. Claude needs a project and a location, from config or the environment, and fails at request time without them. Express-mode API keys don’t work for Claude.
`vertex.apiKey`	`string`	`GOOGLE_VERTEX_API_KEY`	API key for Vertex express mode. Without it, Vertex uses Google Cloud application credentials.
`vertex.googleAuthOptions`	`GoogleAuthOptions`	Application default credentials	Options for `google-auth-library`, such as `credentials` or `keyFile`.
`vertex.headers`	`Record<string, string \| undefined>`, a promise of one, or a function returning either	—	Extra headers. Unlike other providers, Vertex can resolve them per request.
`vertex.toolResultDownloads`	`{ maxBytes?: number }`	`maxBytes`: 7 MiB	Size limit for each remote file downloaded from tool results before it’s sent to Vertex inline.

Custom fetch functions aren’t accepted in config. Pass a pre-built instance instead.

TypeSafe AI evaluation

TypeSafe AI provides evaluation models for POST /v1/evaluate and local frogbot.evaluate(), not chat or embedding models. Set TYPESAFE_AI_API_KEY in the server environment and configure providers: { 'typesafe-ai': {} } in a standalone gateway, or ai: { providers: { 'typesafe-ai': true } } in FrogBot. To use a different server-side secret, set apiKey instead:

ai: {
  providers: {
    'typesafe-ai': { apiKey: process.env.MY_TYPESAFE_KEY }
  }
}

The key stays on the server; authenticate callers in your standalone host or through FrogBot’s authenticated /api/v1/evaluate mount. FrogBot accepts typesafe-ai/jev and typesafe-ai/jev-latest and sends both to TypeSafe as jev-latest. Bare jev is not a documented TypeSafe upstream model ID. An optional built-in models allowlist uses unprefixed IDs, for example models: ['jev-latest']; the alias is canonicalized before checking it. As of September 23, 2026, TypeSafe lists Jev 1.13 at $0.042 per million input tokens, with no charge for output tokens. FrogBot’s USD cost uses these catalog rates as an estimate, not a provider invoice. The upstream jev-latest alias and its pricing can change; check TypeSafe’s model list for current rates.

Vercel AI Gateway

The vercel provider sends requests to Vercel AI Gateway, which serves many model creators through one key. Set AI_GATEWAY_API_KEY in the server environment and configure providers: { vercel: {} } in a standalone gateway, or ai: { providers: { vercel: true } } in FrogBot. AI_GATEWAY_BASE_URL optionally overrides the endpoint. To use a different secret, set apiKey instead. The key is required. A Vercel deployment’s OIDC token does not enable the provider, so a project on Vercel can send its traffic to another provider without AI Gateway turning on. Model IDs have the form vercel/<creator>/<model>, for example vercel/anthropic/claude-sonnet-4.6 or vercel/openai/gpt-5.4-mini. FrogBot accepts the chat models that models.dev lists for Vercel with a price, and uses those prices for usage cost. Other IDs return model_not_found. Image, video, speech, transcription, embedding, and rerank models are not available through this provider. Narrow the list with models, which uses IDs without the vercel/ prefix:

createGateway({
  providers: {
    vercel: { models: ['anthropic/claude-sonnet-4.6', 'openai/gpt-5.4-mini'] }
  }
})

For Anthropic, OpenAI, and Google models, reasoning and cache options are translated as they are for the direct provider. For example, chat reasoning_effort on a Claude model becomes a thinking budget, and cache_control markers become Anthropic cache control. Pass AI Gateway routing options such as order, only, and models in a gateway field on /v1/chat/completions, /v1/messages or /v1/responses, or set providerOptions.gateway in a beforeUpstream hook. /v1/messages also accepts them as providerOptions.gateway in the body, as AI Gateway’s Anthropic API does. They reach AI Gateway unchanged:

{
  "model": "vercel/anthropic/claude-sonnet-4.6",
  "messages": [{ "role": "user", "content": "Hello" }],
  "gateway": { "order": ["bedrock", "anthropic"] }
}

Requests identify FrogBot to AI Gateway with the app attribution headers http-referer: https://www.frogbot.ai and x-title: FrogBot. To use your own values, set them in headers:

createGateway({
  providers: {
    vercel: { headers: { 'http-referer': 'https://example.com', 'x-title': 'Example' } }
  }
})

Bedrock prompt caching

Bedrock Claude models translate Anthropic-style cache_control markers from /v1/chat/completions and /v1/messages into Bedrock cachePoint blocks. Markers are supported on system blocks, messages, and message content parts, including tool results. A request-level marker applies to the last message, caching the conversation prefix through that point. The 5m and 1h TTL values pass through to Bedrock. Other TTL values are omitted without rejecting the request. This translation does not apply to Nova models, tool definitions, or /v1/responses.

OpenAI prompt caching

OpenAI models translate Anthropic-style cache_control markers from /v1/chat/completions and /v1/messages into explicit OpenAI prompt cache breakpoints. Markers are supported on messages and message content parts. A request-level marker applies to the last message. OpenAI retention accepts only in_memory and 24h. Marker TTL values are omitted because OpenAI breakpoints have no TTL equivalent. TTL does not infer retention. This translation does not apply to Azure, custom OpenAI-compatible providers, tool definitions, or /v1/responses.

OpenRouter

OpenRouter gives one API key access to models from many vendors. Set OPENROUTER_API_KEY in the server environment and configure providers: { openrouter: {} } in a standalone gateway, or ai: { providers: { openrouter: true } } in FrogBot. OPENROUTER_BASE_URL optionally overrides the API URL. Model IDs keep OpenRouter’s own <creator>/<model> form after the openrouter/ prefix, including variants such as :free, for example openrouter/anthropic/claude-sonnet-4.6 or openrouter/google/gemma-4-31b-it:free. FrogBot accepts the priced models that models.dev lists for OpenRouter. Dynamic routers without a fixed price, such as openrouter/auto, and models not yet in the catalog return a model-not-found error before any upstream call. Narrow the set with an allowlist, for example models: ['anthropic/claude-sonnet-4.6']; the allowlist applies only to config objects, not to a pre-built instance. Usage cost is estimated from those catalog prices, not read from OpenRouter’s invoice. Requests identify FrogBot to OpenRouter with the X-OpenRouter-Title: FrogBot and HTTP-Referer: https://www.frogbot.ai headers. Override them with appName and appUrl:

providers: {
  openrouter: { appName: 'Acme Support', appUrl: 'https://acme.example' }
}

Reasoning requests from every route become OpenRouter’s reasoning option: reasoning_effort on /v1/chat/completions, thinking on /v1/messages, and reasoning.effort on /v1/responses. Disabled thinking is sent as effort none. Anthropic-style cache_control markers on messages and content parts reach OpenRouter unchanged. On every route, OpenRouter’s own request fields pass through unchanged, for example provider: { order: ['anthropic'], allow_fallbacks: false }, models fallbacks, plugins, or, on /v1/chat/completions, a native reasoning object, which takes priority over reasoning_effort. OpenRouter embeddings are not supported; /v1/embeddings with an openrouter/... model returns an error.

Prefer the built-in openrouter provider over a custom OpenAI-compatible entry pointing at https://openrouter.ai/api/v1. The generic entry drops per-part cache_control, cache-write token counts, OpenRouter’s reasoning format, routing options, and PDFs supplied by URL.

Custom OpenAI-compatible endpoints

Any key that isn’t a built-in provider name is a generic OpenAI-compatible endpoint. baseURL is required — it’s what tells the gateway a key is a deliberate custom endpoint rather than a typo of a built-in name.

providers: {
  ollama: { baseURL: 'http://localhost:11434/v1' },
  'lm-studio': { baseURL: 'http://localhost:1234/v1' },
  vllm: {
    baseURL: 'https://vllm.internal.example/v1',
    apiKey: process.env.VLLM_API_KEY,
    headers: { 'x-deployment': 'production' }
  }
}

A key without baseURL fails validation at startup: providers.ollama.baseURL must be a non-empty string. This is the safety net for a mistyped built-in name (openia instead of openai) — it fails loudly instead of silently becoming an unreachable custom provider.

Custom endpoint config accepts:

Option	Type	Default	Description
`baseURL`	`string`	Required	Base URL of the OpenAI-compatible API, typically ending in `/v1`. Must be non-empty.
`apiKey`	`string`	—	Sent as `Authorization: Bearer <apiKey>`. Omit it for endpoints without authentication.
`headers`	`Record<string, string>`	—	Extra request headers on every call.
`queryParams`	`Record<string, string>`	—	Extra query params appended to every request URL.

The key itself can’t contain /. Custom entries don’t accept models. Custom OpenAI-compatible providers keep their existing routing behavior and are not interpreted as built-in allowlists. In FrogBot’s ai.providers surface, custom entries use baseUrl and a required list of model definitions instead; see Custom provider options. FrogBot router model strings also remain unchanged.

Pre-built instances

Any provider key — built-in or custom — also accepts a pre-built AI SDK provider instance instead of a config object. Use this for a custom fetch, middleware, or anything the shorthand config can’t express.

import { createOpenAI } from '@ai-sdk/openai'
import { createGateway } from '@frogbotai/gateway'

const customOpenAI = createOpenAI({
  fetch: myLoggingFetch
})

const gw = createGateway({
  providers: { openai: customOpenAI }
})

For a known key, the instance is checked against that provider’s real AI SDK type. For a custom key, TypeScript can only confirm it looks like a provider (languageModel present) — the specific instance type isn’t known ahead of time.

Model IDs

Every route dispatches on a canonical <provider>/<model> ID. Built-in model IDs must exist in the configured catalog. Bedrock, Anthropic-on-Bedrock, and Azure additionally resolve friendly aliases to the exact ID the upstream API requires — for example bedrock/claude-4-sonnet resolves to anthropic.claude-sonnet-4-20250514-v1:0 before catalog validation and the AI SDK sees it.

Start here

Basics

Managing Data

Features

Ecosystem

Deployment

Built-in providers

Built-in provider options

TypeSafe AI evaluation

Vercel AI Gateway

Bedrock prompt caching

OpenAI prompt caching

OpenRouter

Custom OpenAI-compatible endpoints

Pre-built instances

Model IDs

​Built-in providers

​Built-in provider options

​TypeSafe AI evaluation

​Vercel AI Gateway

​Bedrock prompt caching

​OpenAI prompt caching

​OpenRouter

​Custom OpenAI-compatible endpoints

​Pre-built instances

​Model IDs

Built-in providers

Built-in provider options

TypeSafe AI evaluation

Vercel AI Gateway

Bedrock prompt caching

OpenAI prompt caching

OpenRouter

Custom OpenAI-compatible endpoints

Pre-built instances

Model IDs