Grok parameter and protocol differences

For Grok models, adjust thinking, sampling and caching using this page. See Chat Completions and Responses for full request and response structures.

Applicable models

Model	Usage notes
grok-4.7	Supports `low/medium/high/xhigh`, default `high`, with thinking always on; Responses returns encrypted thinking by default
grok-4.6	Also supports four levels, default `high`, with thinking always on; request encrypted thinking through `include`
grok-4.5	Supports `low/medium/high`, default `high`, with thinking always on; `xhigh` is treated as `high`

xAI treats Chat as its legacy compatibility endpoint and Responses as the main API; historical Messages compatibility is officially deprecated. This page covers Chat and Responses, not platform discontinuation of other endpoints.

Chat Completions differences

Field	Difference
`reasoning_effort`	Select levels from the table above; `none` cannot disable thinking
`temperature`	Requests allow `0～2`, but actual sampling is capped at `1`; valid values above `1` become `1`. This is not field deprecation
`top_p`	Valid request values go up to `1`, but actual sampling cannot exceed the model default, currently `0.95`; omission uses that default
`presence_penalty/frequency_penalty/stop`	Unsupported on these reasoning models and returns an error; do not reuse these optional fields from the general Reference
`logprobs/top_logprobs`	Unsupported on all three models and silently ignored, unlike the error-returning fields above
`tools[].function.strict`	Tool arguments always follow the supported JSON Schema; `strict` is implicitly `true`. Setting `false` does not enable non-strict generation
Thinking history	Chat has no field carrying Responses-style encrypted thinking; use Responses for encrypted thinking across turns

Responses differences

Field	Difference
`reasoning.effort`	Use the model's levels above, default `high`; thinking cannot be disabled
`temperature/top_p`	The same sampling caps apply. Responses may echo original request values rather than actual capped sampling settings
Penalties and Logprobs	Same limits as Chat: penalties are unsupported; Logprobs parameters are ignored. Switching to Responses does not enable them
`max_output_tokens`	Limits visible answers only, excluding thinking and function-call tokens; default `128000`. Do not interpret it as OpenAI's combined budget
`include`	grok-4.6 and grok-4.5 request encrypted thinking through `reasoning.encrypted_content`; grok-4.7 returns it even without the request
`store/previous_response_id`	`store` defaults to `true`; vendor responses last `30` days and can be continued by ID. With `store: false`, manually replay complete history and all output items, not just final text; preserve encrypted thinking unchanged
`tools[].strict`	Function tools also use implicit strict mode; function parameters and structured output use xAI's supported JSON Schema subset
JSON Schema	`additionalProperties` defaults to `false`; explicitly set `true` to allow extra properties. `$ref/$defs` supports noncircular references only, not recursive schemas
Server-side tools	xAI Responses uses types such as `web_search`, `x_search` and `code_interpreter`; the latter corresponds to native xAI SDK `code_execution`. Do not mix these names

Cache differences

Field	Difference
Automatic prefix caching	Identical input prefixes are reused automatically; put static instructions first, append later content, and do not edit or reorder history
`x-grok-conv-id`	Stable conversation identifier in a Chat Header, helping route the conversation to servers holding its cache; not a model parameter
`prompt_cache_key`	The corresponding identifier in the Responses body; keep it stable across turns and replay encrypted thinking unchanged. Do not add other vendors' TTL or explicit-breakpoint settings
Cache usage	Chat: inspect `usage.prompt_tokens_details.cached_tokens`; Responses: `usage.input_tokens_details.cached_tokens`. Cache may be evicted; a conversation identifier does not guarantee a hit

Media input differences

Input	Protocol content block	Difference
Image	`image_url`	Chat `image_url.url` accepts public URLs or Base64 Data URLs; JPEG / PNG, maximum `20 MiB` per image
Image	`input_image`	Responses `image_url` directly contains a public URL or Base64 Data URL, with the same limits; not Chat's nested object

All three output text. detail supports auto/low/high for image processing detail, not image generation.

Request examples

Replace MODEL_ID with the full platform model ID and set a stable identifier per conversation. Responses uses manual history management; replay full output and encrypted thinking in later turns.

curl 'https://api.agtcloud.ai/v1/chat/completions' \
  -H 'Authorization: Bearer your-api-key' \
  -H 'Content-Type: application/json' \
  -H 'x-grok-conv-id: YOUR_CONVERSATION_ID' \
  --data '{
    "model": "MODEL_ID",
    "messages": [{"role": "user", "content": "Explain how to troubleshoot an API timeout in three points"}],
    "reasoning_effort": "high",
    "max_tokens": 4096
  }'