Anthropic Message API (Subscription)

Model ID. Use an authorized ID returned by GET /v1/models.

Maximum output tokens. Required — send a positive integer. Omitting it currently does not error on this channel, but the effective default is unknown and the behaviour is not guaranteed.

User / assistant messages. Text, images, search results, tool calls, and tool results were verified. Document processing requires a real valid PDF to verify end-to-end.

Values: "user" "assistant"

Values: "text" "image" "document" "search_result" "tool_use" "tool_result"

Top-level system prompt; do not put it in messages.

Values: "text"

Whether to return SSE streaming events. In the 2026-08-25 HopsAPI retest true, false, and omission all eventually succeeded, but non-streaming requests showed intermittent 424 the same day; for stability prefer stream: true.

Array of custom stop strings.

Extended thinking; enabled, disabled, and adaptive were verified.

Values: "enabled" "disabled" "adaptive"

Required when type is enabled; minimum 1024 (below 1024 returns 400 in testing). The official spec requires it to be less than max_tokens, but this channel does not enforce that: budget_tokens >= max_tokens still returns 200, with output truncated at max_tokens.

Values: "summarized" "omitted"

Client tool definitions. Strict schemas verified; the two-turn tool_result loop is verified on Claude Opus 5, while Claude Sonnet 5 has a known issue of returning 200 without using the tool result (tested 2026-08-18).

JSON Schema for tool input. For an object schema with strict=true, include additionalProperties=false.

Tool selection object. A string such as auto is rejected; use {type: auto}, {type: any}, {type: tool, name: ...}, or {type: none}.

Values: "auto" "any" "tool" "none"

Not a top-level parameter — sending it at the root has no effect. cache_control is a marker placed inside the block you want cached: a messages[].content[] block, a system block, or a tools[] entry. Prompt caching works on this channel: once a block is marked, repeated requests report cache_read_input_tokens in the response.

Values: "ephemeral"

Values: "5m" "1h"

Context management; the tested shape was successful.

Anonymous request metadata; user_id was verified.

JSON Schema structured output through output_config.format.

Values: "json_schema"

Service tier; auto and standard_only were verified.

Values: "auto" "standard_only"

Unsupported: standard and fast return Extra inputs are not permitted.

Unsupported: returns Extra inputs are not permitted.

Unsupported: returns Extra inputs are not permitted.

Unsupported on the current path.

Unsupported: returns Extra inputs are not permitted.

Account restricted: unsupported for this organization type.

Deprecated: use output_config.format.

Unsupported for Claude Opus 5 on this channel: the upstream rejects the request with a deprecated error for this model. The parameter itself is still current in the official Anthropic API — the restriction is the model, not the protocol. Official use: sample from only the top K options for each subsequent token.

Unsupported for Claude Opus 5 on this channel: the upstream rejects the request with a deprecated error for this model. The parameter itself is still current in the official Anthropic API — the restriction is the model, not the protocol. Official use: nucleus sampling, 0 to 1.

This channel rejects the parameter as deprecated for this model: temperature: 0 returns 400 (message: temperature is deprecated for this model.); only 1 (the default) passes. Omitting the field uses the model default.

Shape pending: a request containing expires_at was rejected, which does not prove every shape is unsupported.