Reasoning tokens used up an output cap on GPT-5.5 Pro
- Date
- 3 August 2026
- Task type
- API behaviour
- Status
- Complete
Models tested
| Model or tool | Provider | Open? | How run |
|---|---|---|---|
| GPT-5.5 Pro (gpt-5.5-pro-2026-04-23, Responses API) | OpenAI | Closed | API |
Purpose
Record what happens when a request to a reasoning model sets a maximum output length that the model's reasoning alone can use up.
Context
Two requests to the Responses API in a cost-controlled consultation. Costs are at the provider's list rates of $30 per million input tokens and $180 per million output tokens.
Method
The first request used medium reasoning and max_output_tokens of 6,000. A second request on a revised problem used high reasoning and no output cap. Status, token usage, time and visible text were recorded from the response.
Results
| Request | Reasoning setting | Output cap | Status | Time | Input tokens | Output tokens | Reasoning tokens | Visible answer tokens | Approximate cost |
|---|---|---|---|---|---|---|---|---|---|
| Capped (streaming retry) | medium | 6,000 | incomplete | 101.25 s | 1,718 | 6,000 | 6,000 | 0 | $1.13 |
| Uncapped | high | none | completed | 541.92 s | 1,940 | 22,828 | 19,793 | 3,035 | $4.17 |
Caveats
- Two requests only (n=1 each), and they used different reasoning settings and different problems.
- The first non-streaming attempt ended after about 40 seconds with a disconnection; its usage and cost could not be established, so the capped cost excludes it.
What this suggests
The output cap covered reasoning and visible text together, so a cap that looks generous for the answer can be spent before any answer is produced. The uncapped run used more than three times the capped limit on reasoning alone.
Source
entries/2026-08-03-reasoning-tokens-output-cap.md in the GitHub repository (Markdown with YAML frontmatter).