Claude parameter and protocol differences

Claude natively uses Messages. For Chat Completions, consult the Anthropic compatibility-layer limits below; platform conversion depends on the selected model endpoint.

Applicable models

Model	Thinking defaults and controls
claude-haiku-5-5	Enabled by default at effort `medium`; `disabled` can disable it only at `low/medium/high`
claude-sonnet-5-5	Enabled by default at effort `high`; `disabled` is rejected. At `low/medium/high`, use `between_tools` to disable thinking before answers
claude-opus-5-5	Enabled by default at effort `medium`; cannot be disabled
claude-fable-5-1	Enabled by default at effort `high`; cannot be disabled
claude-opus-5	Enabled by default at effort `high`; `disabled` can disable it only at `low/medium/high`
claude-sonnet-5	Enabled by default at effort `high`; disable with `disabled`
claude-fable-5	Enabled by default at effort `high`; cannot be disabled
claude-opus-4-8	Disabled by default, with effort `high`; enable with `adaptive`

Messages differences

Field	Difference
`thinking.type`	All eight support `adaptive` and reject the old `enabled` and `budget_tokens`. See the table above for disabling support; `disabled` does not apply to every model
`between_tools`	Sonnet 5.5 only; disables thinking before answers, while progress between tools still returns in `thinking` blocks. The object may contain only `type`, not `display` or `budget_tokens`
`output_config.effort`	All eight support `low/medium/high/xhigh/max`, with defaults above; `adaptive` is a thinking type, not effort. `max_tokens` limits thinking and the final answer together
`thinking.display`	All eight default to `omitted` when thinking is enabled, returning empty thinking text and `signature`. Set `summarized` for a readable summary. Empty text does not mean no thinking occurred
`tool_choice`	Opus 5.5, Sonnet 5.5 and Fable 5.1 reject `any` and a named `tool`; use `auto/none`. The other five allow forced calls, which skip thinking on that call. Fable 5 does not share Fable 5.1's restriction
Thinking history and signatures	After tool calls, replay previous `thinking/redacted_thinking` and tool blocks completely, including empty thinking text. Do not edit, reorder or rebuild signatures. Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1 also validate original system prompts, tools and earlier history; changed history cannot directly reuse old thinking blocks
Assistant prefilling	All eight reject a final unfinished assistant message with `400`; constrain output through instructions or `output_config.format`
`temperature/top_p/top_k`	Omit from examples. `2023-06-01` is also subject to newer-model sampling restrictions; under `2026-10-21`, supplying any of these fields for any model returns `400`. Switching versions does not restore custom sampling
Computer tool version	In the Claude API, Haiku 5.5, Sonnet 5.5 and Opus 5.5 use `computer_toolset_20260801`; old Computer tool types no longer apply. See official guidance for other cloud platforms

Official Chat compatibility differences

These rules describe Anthropic's OpenAI SDK compatibility layer, distinct from native Messages capabilities; platform channels may use different conversions.

Field	Difference
`n`	Only accepts `1`
`reasoning_effort`	Ignored; use Messages `output_config.effort` to control effort or access thinking
`response_format`, `tools[].function.strict`	Ignored; these do not enable native structured output or strict tool validation
`presence_penalty/frequency_penalty/logit_bias/seed`	Ignored; no corresponding sampling control
`logprobs/top_logprobs`	Ignored; no token probabilities are returned
`system/developer`	Merged into the opening system prompt with newline separators; original positions within the conversation are not retained
`tool_choice`	`required` and named functions are subject to the three-model forced-call restriction above
Caching and other settings	No native prompt-cache configuration; fields such as `store/service_tier/metadata/prediction/user/audio/modalities` are ignored and do not enable their namesake capabilities

Cache differences

Field	Difference
`cache_control`	Use Messages `type: ephemeral`; choose breakpoints automatically at request level or explicitly on content blocks, up to four breakpoints
`ttl`	Default: `5m`; optional: `1h`; distinct from Chat's `prompt_cache_options`
Minimum cacheable prefix	Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1, Opus 5 and Fable 5 require `512` input tokens; Sonnet 5 and Opus 4.8 require `1024`. Prefixes must match; below the threshold, no cache is written
Thinking blocks	Cannot mark `thinking` blocks directly; replayed thinking can be cached as part of preceding assistant content
Cache usage	Inspect `usage.cache_creation_input_tokens` and `usage.cache_read_input_tokens` for writes and reads; setting cache options alone does not prove a hit

Media input differences

Input	Protocol content block	Difference
Image	Messages `image`	Use URL or Base64 within `source`, not Chat's `image_url` block
Document	Messages `document`	Use the native document structure in user messages, not assistant messages
Image	Chat `image_url`	The official compatibility layer reads `image_url.url` and ignores `detail`
File	Chat `file`	The official compatibility layer ignores this input; it does not provide native document capabilities
Audio	Chat `input_audio`	The official compatibility layer ignores this input; it does not provide native audio capabilities

Request examples

Replace MODEL_ID with the full platform model ID. The Messages example explicitly enables adaptive thinking and a summary; Chat demonstrates basic text only.

curl 'https://api.agtcloud.cn/v1/messages' \
  -H 'x-api-key: your-api-key' \
  -H 'anthropic-version: 2023-06-01' \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "MODEL_ID",
    "messages": [{"role": "user", "content": "Explain how to troubleshoot an API timeout in three points"}],
    "max_tokens": 4096,
    "thinking": {"type": "adaptive", "display": "summarized"},
    "output_config": {"effort": "medium"}
  }'