Anthropic

Configure a deployment

Send this body to POST /admin/deployments with an operator credential. Replace the example credentials and choose a model available to your provider account. See Creating deployments for the shared configuration fields.

{ "publicModel": "claude", "adapterKey": "anthropic", "upstreamModel": "claude-sonnet-5", "credentials": { "apiKey": "sk-ant-..." } }

Behavior and limitations

• Credentials: apiKey. version (the anthropic-version header) defaults to 2023-06-01 and rarely needs overriding.
• Default transport: Messages API for text.
• Reasoning: either anthropic_adaptive (the model picks its own depth; the request only carries an output_config.effort label) or anthropic_budget (thinking: { budget_tokens }, with a provider-side default budget per canonical level if the catalog doesn't override it) — see Reasoning. Models that expose both xhigh and max preserve them as distinct tiers; models such as Opus/Sonnet 4.6 declare max without inventing xhigh support.
• Signed thinking round-trips on /v1/responses. When the request includes reasoning.encrypted_content, each turn's content (thinking, text and tool calls, in order) is sealed into a reasoning item that standard clients replay, and the next request sends the turn back exactly as Claude returned it. On /v1/messages the blocks stay native, in Claude's order. See Provider-specific fields.
• Claude Opus 5 enables adaptive thinking upstream when thinking is omitted. The adapter sends thinking: { "type": "disabled" } when canonical reasoning resolves to none, preserving the gateway's provider-independent off semantics.
• claude-opus-5-fast and claude-opus-4-8-fast are gateway aliases for Anthropic's research-preview fast mode. The adapter selects the base model, sends speed: "fast", and adds the required beta header. Anthropic account access is still required.
• Native /v1/messages requests pass through with minimal translation; /v1/chat/completions and /v1/responses requests are rendered onto the same canonical reasoning/tool/format machinery as every other provider.
• Prompt caching is automatic on /v1/chat/completions and /v1/responses. The adapter sends Anthropic's top-level cache_control: { "type": "ephemeral" } (5-minute TTL), so repeated prefixes are cached as they are on OpenAI and Gemini and reported in cached_tokens. To control caching yourself, use /v1/messages: those requests are sent with exactly the cache_control you write, or none.
• Models that declare capabilities.strictTools: true preserve Anthropic's native top-level strict: true tool flag. Native output_config.format is schema-adherent and is treated as strict structured output by the gateway.
• Anthropic compiles strict tool schemas and json_schema response formats from a subset of JSON Schema and rejects constraints outside it (maxItems, minLength, minimum, …) that OpenAI's strict mode accepts. For /v1/chat/completions and /v1/responses requests the adapter keeps the supported keywords and moves the rest into the field's description, as Anthropic's SDKs do; clients keep validating against their original schema. Native /v1/messages tools are sent as written.
• Catalog: src/adapters/anthropic/catalog.json.

Next steps

• Messages — the native Anthropic-compatible contract.
• Reasoning — anthropic_adaptive / anthropic_budget in detail.