gpt-6.1-sol and gpt-6-astra use low/medium/high/xhigh/max; the other six also support none. None of these eight lists minimal as available
tools
See model limits above; even the five models supporting Chat tools with thinking disabled only call client-side tools. Use Responses for OpenAI-hosted tools
temperature, top_p
For GPT-6 / 6.1, omit these unless effort is none; the two models that cannot disable thinking must omit them entirely. A model-specific restriction, not general field deprecation
logprobs, top_logprobs
Omit for GPT-6 / 6.1 with thinking enabled; do not reuse the general Reference's Logprobs example
max_completion_tokens
Thinking and the final answer share the budget. Reserve output tokens for both, not just the visible answer
Choose available levels by model as in the Chat table; Responses does not remove a model's restriction on none
reasoning.mode
GPT-5.6 and GPT-6 / 6.1 support standard/pro, defaulting to standard. pro increases model work independently of effort, without changing model ID
reasoning.context
The three GPT-5.6 models default to all_turns; gpt-6.1-sol also supports it. auto uses the model's default policy. Preserve actual response history for reuse; incompatible thinking is not carried across model families
configuration_update
GPT-6 / 6.1 in standard single-agent mode accepts this item in input to adjust later effort. Keep top-level reasoning.effort unchanged to preserve the original prefix cache. Unavailable for GPT-5.6 or mode pro
temperature, top_p, top_logprobs
For GPT-6 / 6.1, omit unless effort is none, and remove message.output_text.logprobs from include
Historical reasoning items
Continue with previous_response_id or manually replay complete input and output items. With store: false, preserve the returned encrypted reasoning items, not summaries as original reasoning
All eight support implicit/explicit; the former automatically selects cacheable message boundaries, the latter uses explicit breakpoints. Minimum cacheable prefix: 1024 visible input tokens, with an exact prefix match still required
prompt_cache_breakpoint
Mark explicit breakpoints on applicable input blocks. Changing history, tools or output configuration can change the prefix; using the same model name does not guarantee a hit
prompt_cache_options.ttl
Default: 30m; Responses also supports 12h, a minimum retention time, not a deletion deadline. Do not apply Responses 12h to Chat
prompt_cache_key
Optional cache-attribution identifier for GPT-5.6 and later; not required to enable caching or achieve a hit
Cache usage
Inspect token-detail cached_tokens and cache_write_tokens for cache reads and writes; enabling caching in a request does not prove a hit
Replace MODEL_ID with the full platform model ID. Chat has no tools in this example; Responses demonstrates structured output. Both use effort low, supported by all eight models.
curl 'https://api.agtcloud.ai/v1/chat/completions' \ -H 'Authorization: Bearer your-api-key' \ -H 'Content-Type: application/json' \ --data '{ "model": "MODEL_ID", "messages": [{"role": "user", "content": "Explain how to troubleshoot an API timeout in three points"}], "reasoning_effort": "low", "max_completion_tokens": 4096 }'