Comet API

Comet API is supported as an Anthropic-compatible provider via the dedicated cometapi credential type. By default, requests are converted from OpenAI Chat Completions or Responses API format to Anthropic Messages API and sent to /v1/messages. Comet also exposes an OpenAI-compatible /v1/chat/completions endpoint for the same models; set openai_proto: true on a credential to use that wire protocol instead (see OpenAI Protocol Mode). For Google/Gemini models Comet additionally exposes a Google GenAI-compatible /v1beta/models/{model}:generateContent endpoint; set google_proto: true to use it (see Google Protocol Mode).

Configuration

credentials:
  - name: "comet_anthropic"
    type: "cometapi"
    api_key: "os.environ/COMET_API_KEY"
    base_url: "https://api.cometapi.com/v1"
    rpm: 60
    tpm: -1

OpenAI Protocol Mode

Set openai_proto: true to switch a cometapi credential from Comet's Anthropic-compatible /v1/messages endpoint to its OpenAI-compatible /v1/chat/completions endpoint. Requests and responses are then passed through unconverted (the same code path used for type: openai), instead of being translated to/from Anthropic Messages format:

credentials:
  - name: "comet_openai"
    type: "cometapi"
    api_key: "os.environ/COMET_API_KEY"
    base_url: "https://api.cometapi.com/v1"
    openai_proto: true
    rpm: 60
    tpm: -1

openai_proto is only valid on type: cometapi credentials — setting it on any other provider type fails config validation. It is also mutually exclusive with google_proto.

With openai_proto: true:

The upstream URL is built from base_url + the incoming request path (typically /v1/chat/completions), not /v1/messages.
Auth is a plain Authorization: Bearer <api_key> header; auth_type and the anthropic-version header are not used.
The Claude Model Aliases and Prompt Caching sections below (Anthropic cache_control format, the anthropic-beta header) do not apply — the request body is never translated to Anthropic's shape.
Error Masking still applies, since it is keyed on the credential's identity (cometapi), not its wire protocol.

Use separate credentials (e.g. comet_anthropic and comet_openai) if you want some models routed via Anthropic protocol and others via OpenAI protocol at the same time.

Google Protocol Mode

Set google_proto: true to switch a cometapi credential from Comet's Anthropic-compatible /v1/messages endpoint to its Google GenAI-compatible /v1beta/models/{model}:generateContent (and :streamGenerateContent?alt=sse) endpoint. This reuses the exact same request/response path as type: gemini (Google AI Studio): incoming OpenAI Chat Completions / Responses / embeddings requests are translated to Google's contents / generationConfig shape and the Gemini response is translated back to OpenAI format.

credentials:
  - name: "comet_google"
    type: "cometapi"
    api_key: "os.environ/COMET_API_KEY"
    base_url: "https://api.cometapi.com"
    google_proto: true
    rpm: 60
    tpm: -1

models:
  - name: "gemini-3.1-flash"
    credential: comet_google
  - name: "gemini-3.1-flash-image-preview"
    credential: comet_google

Note the base_url has no /v1 suffix — the /v1beta/models/... path is appended by the router (matching the google-genai SDK's base_url="https://api.cometapi.com", api_version="v1beta").

With google_proto: true:

The upstream URL is base_url + /v1beta/models/<model>:generateContent (streaming: :streamGenerateContent?alt=sse), not /v1/messages.
Auth travels as the x-goog-api-key: <api_key> header (the CometAPI key); auth_type and the anthropic-version header are not used.
Image generation (responseModalities, imageConfig), streaming, and /v1/responses all work exactly as they do for type: gemini.
The Claude Model Aliases and Prompt Caching sections below (Anthropic cache_control format, the anthropic-beta header) do not apply — the request body is never translated to Anthropic's shape.
Error Masking still applies, since it is keyed on the credential's identity (cometapi), not its wire protocol.

google_proto is only valid on type: cometapi credentials and is mutually exclusive with openai_proto. Use separate credentials (e.g. comet_anthropic, comet_openai, comet_google) to route different models through different protocols at the same time.

Claude Model Aliases

Use public name values for clients, and map them to Comet model IDs with model where the provider names differ:

models:
  - name: "anthropic/claude-haiku-4.5"
    model: "claude-haiku-4-5-20251001"
    credential: comet_anthropic
  - name: "claude-haiku-4.5"
    model: "claude-haiku-4-5-20251001"
    credential: comet_anthropic
  - name: "claude-haiku-4-5-20251001"
    credential: comet_anthropic

  - name: "anthropic/claude-opus-4.1"
    model: "claude-opus-4-1-20250805"
    credential: comet_anthropic
  - name: "claude-opus-4.1"
    model: "claude-opus-4-1-20250805"
    credential: comet_anthropic
  - name: "claude-opus-4-1-20250805"
    credential: comet_anthropic

  - name: "anthropic/claude-opus-4.5"
    model: "claude-opus-4-5-20251101"
    credential: comet_anthropic
  - name: "claude-opus-4.5"
    model: "claude-opus-4-5-20251101"
    credential: comet_anthropic
  - name: "claude-opus-4-5-20251101"
    credential: comet_anthropic

  - name: "anthropic/claude-opus-4.6"
    model: "claude-opus-4-6"
    credential: comet_anthropic
  - name: "claude-opus-4-6"
    credential: comet_anthropic

  - name: "anthropic/claude-opus-4.7"
    model: "claude-opus-4-7"
    credential: comet_anthropic
  - name: "claude-opus-4-7"
    credential: comet_anthropic

  - name: "anthropic/claude-sonnet-4"
    model: "claude-sonnet-4-20250514"
    credential: comet_anthropic
  - name: "claude-sonnet-4"
    model: "claude-sonnet-4-20250514"
    credential: comet_anthropic

  - name: "anthropic/claude-sonnet-4.5"
    model: "claude-sonnet-4-5-20250929"
    credential: comet_anthropic
  - name: "claude-sonnet-4.5"
    model: "claude-sonnet-4-5-20250929"
    credential: comet_anthropic
  - name: "claude-sonnet-4-5-20250929"
    credential: comet_anthropic

  - name: "anthropic/claude-sonnet-4.6"
    model: "claude-sonnet-4-6"
    credential: comet_anthropic
  - name: "claude-sonnet-4-6"
    credential: comet_anthropic

Prompt Caching

cometapi uses the Anthropic Messages cache-control format:

{"cache_control": {"type": "ephemeral"}}

When session-sticky routing is active, the router can automatically inject cache markers for stable conversation history. Cache accounting is preserved in OpenAI responses:

Comet/Anthropic usage field	OpenAI-compatible response field
`cache_read_input_tokens`	`prompt_tokens_details.cached_tokens`
`cache_creation_input_tokens`	`prompt_tokens_details.cache_creation_tokens`

For claude-sonnet-4.5, prefer the dated Comet model claude-sonnet-4-5-20250929 for more stable prompt-cache behavior.

Error Masking

Comet upstream error bodies are masked before being returned to clients or logged. The HTTP status is preserved, but the response body is replaced with a neutral OpenAI-compatible error object.