LLM Parameters

Bold = name differs from the OpenAI Chat Completions spec (gpt-4o = baseline). Dashes = not supported.

1. Sampling / decoding

Model (API)	n	max_tokens	temperature	top_p	top_k	stop	seed
gpt-4o (Chat Completions)	n	max_completion_tokens ¹	temperature ²	top_p ³	—	stop	seed
gpt-5.2 (Responses)	— ⁴	max_output_tokens	temperature ²	top_p ³	—	—	—
claude-opus-4.8 (Messages) ⁵	—	max_tokens ⁶	temperature ⁷	top_p	top_k	stop_sequences	—
gemini-3.5-flash (generateContent) ⁸	candidateCount ⁹	maxOutputTokens	temperature ¹⁰	topP ¹¹	topK	stopSequences	seed
mistral-large (Chat)	n	max_tokens	temperature	top_p ³	—	stop	random_seed
command-a (Cohere v2)	—	max_tokens	temperature	p	k	stop_sequences	seed
llama / OSS (vLLM)	n	max_tokens	temperature	top_p	top_k ¹²	stop	seed

1. max_tokens still works on non-reasoning models.
2. Default 1, range 0–2.
3. Default 1.
4. Not supported — send multiple requests.
5. Identical shape for sonnet-5 and all Claude 4.x.
6. Required — only API where it isn't optional.
7. Default 1, range 0–1.
8. Identical shape for gemini-2.5-pro and all Gemini 2.x/3.x.
9. Alternative candidates per input.
10. Default 1.0.
11. Default 0.95.
12. vLLM extension, not in the OpenAI spec.

2. Structural

Model (API)	system prompt	stream	penalties	logprobs	tools	response_format
gpt-4o	messages[role="system"]	stream	frequency_penalty, presence_penalty	logprobs + top_logprobs	tools, tool_choice	response_format
gpt-5.2	instructions / role developer	stream	—	top_logprobs	tools, tool_choice	text.format
claude-opus-4.8	system (top-level)	stream	—	—	tools, tool_choice	— ¹
gemini-3.5-flash	systemInstruction	separate endpoint ²	frequencyPenalty, presencePenalty	responseLogprobs + logprobs	tools, toolConfig	responseMimeType + responseSchema
mistral-large	messages[role="system"]	stream	frequency_penalty, presence_penalty	—	tools, tool_choice	response_format
command-a	messages[role="system"]	stream	frequency_penalty, presence_penalty	logprobs	tools, tool_choice	response_format
llama / OSS	messages[role="system"]	stream	both	logprobs	tools	guided_json

1. No JSON mode — use a tool schema or prefill.
2. streamGenerateContent, not a flag.

3. Reasoning / thinking

Baseline for bolding here is gpt-5.2 — gpt-4o has no reasoning params.

Model (API)	enable	depth control	token budget	visibility	constraints
gpt-4o	n/a	—	—	—	—
gpt-5.2	on by default	reasoning.effort ¹	max_output_tokens (shared)	reasoning.summary	reasoning tokens billed as output
claude-opus-4.8	thinking object ²	output_config.effort	thinking.budget_tokens ³	thinking blocks; display: "omitted"	no temperature / top_k / forced tools; top_p only 0.95–1 ⁴
gemini-3.5-flash	thinkingConfig	thinkingLevel	thinkingBudget ⁵	includeThoughts	budget 0 disables on Flash; not disableable on Pro
mistral-large	n/a ⁶	—	—	—	—
command-a	thinking object ⁷	—	thinking.token_budget	thinking block	—
llama / OSS	chat_template_kwargs ⁸	reasoning_effort ⁹	max_tokens	reasoning_content	varies per template

1. none / low / medium / high / xhigh.
2. {"type": "adaptive"} on 4.6+; omitting it defaults to adaptive.
3. Min 1024, must be < max_tokens. Deprecated on 4.6+ in favor of effort.
4. The gotcha — table 1 partly stops applying once thinking is on.
5. -1 = model decides.
6. Model choice (magistral), not a param.
7. On command-a-reasoning.
8. e.g. {"enable_thinking": true}.
9. Some vLLM builds only.

Takeaways

• Google camelCases everything and moves sampling into a nested generationConfig; streaming is a different endpoint.
• Anthropic renames a few and drops n, seed, logprobs, and penalties entirely; max_tokens is mandatory.
• Mistral, Cohere, vLLM are ~OpenAI clones — Cohere's p/k and Mistral's random_seed are the only real trip-ups.
• OpenAI's own two APIs disagree (Chat Completions vs Responses) more than some cross-vendor pairs do.
• Reasoning is converging on an effort knob; explicit token budgets are on the way out.