Messages

Send messages using the Anthropic API format. The selected deployment can use Anthropic or another provider that supports the requested capabilities.

Example request

Set GATEWAY to the gateway origin and API_KEY to your virtual key. Replace the example public model with one from Model discovery.

curl -X POST "$GATEWAY/v1/messages" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Say hi" }]
  }'

The native Anthropic wire shape — stop_sequences instead of stop, system as a top-level field or message, output_config.format instead of response_format — rendered onto the same canonical request as Chat completions and Responses. You can request any provider through this contract, not just Anthropic models; the reverse also works (Anthropic models are reachable through Chat Completions and Responses too).

Messages document blocks support URL, base64, and provider file sources. They stay native on the Anthropic transport and use the same portable routing policy elsewhere. See Content inputs.

The router prefers the native Messages transport when it is available. top_k is canonicalized and translated to providers that expose the same control. Request metadata is native-only and requires a Messages-capable deployment instead of being silently dropped.

Prompt-caching breakpoints

cache_control: { "type": "ephemeral" } on a system block, content block, or tool is preserved end-to-end as an opaque pass-through — the canonical core doesn't interpret it, only the Anthropic adapter emits it. Because dropping a cache breakpoint changes billing and latency semantics, requests that contain one require a native Messages deployment instead of silently degrading on another wire.

Structured output

output_config.format maps to the same canonical responseFormat as response_format elsewhere — json_schema is native strict structured output and requests a schema-adherent response on models that declare structuredOutputs: true. Anthropic tool definitions preserve strict: true end to end.

Reasoning

Anthropic's thinking block is resolved from the same canonical reasoning.effort/summary as every other contract. Two distinct kinds exist in the catalog — anthropic_adaptive (the model picks its own depth, driven by output_config.effort) and anthropic_budget (an explicit budget_tokens per level) — see Reasoning for the exact wire shape of each and which models use which. For adaptive models, output_config.effort: "max" stays distinct from "xhigh" end to end.

Signed thinking and redacted thinking blocks are opaque provider state. The gateway preserves and replays the complete blocks, including signatures and redacted data, but does not fabricate signed blocks when a different provider only returns a visible reasoning summary.

Error shape

Errors follow Anthropic's own envelope ({"type": "error", "error": {"type": "...", "message": "..."}}) on this endpoint specifically, even though the gateway's internal classification is identical across all three text contracts — see Troubleshooting for the class-to-type mapping.

Streaming

Same SSE mechanics as every other endpoint, rendered into Anthropic's own event names (message_start, content_block_delta, ...). See Streaming.

Next steps

• Reasoning — anthropic_adaptive vs. anthropic_budget in depth.
• Chat completions and Responses — the other two text contracts.
• Providers → Anthropic — credentials and adapter specifics.