Reasoning effort

What the five effort levels trade off, how reasoning tokens are billed, and how to choose.

GPT-6 Astra and the GPT-5.6 family are reasoning models: before producing visible output they can spend tokens thinking. The reasoning_effort parameter (Chat Completions) or reasoning.effort (Responses API) controls how much. It affects both latency and output-token cost.

The trade-off

Higher effort buys depth — more hypotheses considered, more self-checking — at the price of latency and output-token spend. Reasoning tokens are generated before the first visible token, so effort directly moves time-to-first-token as well as total cost.

Effort	Behavior	Typical fit
`none`	Few or no reasoning tokens; fastest time-to-first-token	Extraction, classification, reformatting, simple lookups
`low`	Brief reasoning	Summarization, straightforward Q&A, routine code edits
`medium` (default)	Balanced	General coding, analysis, multi-step tool use
`high`	Extended reasoning; slow and expensive	Hard debugging, mathematical derivation, long-horizon planning
`xhigh`	Maximum reasoning depth; slowest and most expensive	Frontier-difficulty problems where correctness dominates all other costs

Rough intuition on spend: a request that produces a 300-token answer might generate ~0 reasoning tokens at none, a few hundred at medium, and several thousand at high or xhigh on a genuinely hard problem. Check usage.completion_tokens_details.reasoning_tokens (Chat Completions) or usage.output_tokens_details.reasoning_tokens (Responses) on your real traffic before deciding.

Billing

Reasoning tokens are billed as output tokens, at the output rate (at the selected model's output rate). This is the industry standard — OpenAI, Azure, and every reseller of reasoning models bill this way — and Draftworks is no exception. They are included in completion_tokens / output_tokens and count against max_completion_tokens / max_output_tokens. The receipt's output_tokens field includes them, so billed amounts are independently checkable.

Practical consequence: when setting max_completion_tokens for a high-effort request, budget for reasoning. A cap of 1,000 tokens on a high-effort request can be consumed entirely by reasoning, yielding a truncated or empty visible answer with finish_reason: "length".

Choosing a level

Start at medium and move down. Most production traffic is simpler than it looks; if low holds your quality bar, it is meaningfully faster and cheaper.
Reserve high for requests where a wrong answer costs more than a few extra seconds and cents, and xhigh for the rare problems where you would happily wait minutes.
Use lower effort for latency-sensitive tasks when the selected model supports it. Astra supports low, medium, high, xhigh, and max; it does not support none.
Effort is per request, not per key or per account. Route each call site independently — one application can use none for routing and high for the final synthesis in the same flow.

Setting it

Chat Completions

Python

completion = client.chat.completions.create(
    model="gpt-5.6-sol",
    reasoning_effort="high",
    max_completion_tokens=16384,  # leave room for reasoning
    messages=[{"role": "user", "content": "Find the bug in this concurrent queue: ..."}],
)
print(completion.usage.completion_tokens_details.reasoning_tokens)

Responses API

TypeScript

const response = await client.responses.create({
  model: "gpt-5.6-sol",
  reasoning: { effort: "none" },
  input: "Classify this ticket as billing, bug, or feature-request: ...",
});
console.log(response.usage?.output_tokens_details?.reasoning_tokens);

The Responses API additionally accepts reasoning.summary (auto, concise, detailed) to receive a natural-language summary of the reasoning in the output items. The raw chain of thought is never returned by either API.

Other providers

Claude supports model-specific thinking options through the Bedrock adapter. GLM 5.3 always reasons and accepts low, high, or max; disabling thinking is not supported. Check the selected model before carrying parameters over from another provider.