Reasoning Effort Levels and Defaults by Model
Published reasoning-effort levels and defaults, every listed model
| GPT-6 Astra | low · medium · high · xhigh · max — default: not marked on its page |
|---|---|
| GPT-6.1 Sol | low · medium · high · xhigh · max — default: medium |
| GPT-6.1 Terra | none · low · medium · high · xhigh · max — default: medium |
| GPT-6 Sol | none · low · medium · high · xhigh · max — default: medium |
| GPT-6 Luna | none · low · medium · high · xhigh · max — default: medium |
| GPT-5.6 Sol | none · low · medium · high · xhigh · max — default: medium |
| GPT-5.6 Terra | none · low · medium · high · xhigh · max — default: medium |
| GPT-5.6 Luna | none · low · medium · high · xhigh · max — default: medium |
| GPT-5.5 | none · low · medium · high · xhigh — default: medium |
| GPT-5.5 Pro | medium · high · xhigh — default: high |
| GPT-5.4 | none · low · medium · high · xhigh — default: none |
| GPT-5.4 Mini | none · low · medium · high · xhigh — default: none |
| GPT-5.4 Nano | none · low · medium · high · xhigh — default: none |
| GPT-5.4 Pro | medium · high · xhigh — default: medium |
| GPT-5.3 Codex | low · medium · high · xhigh — default: not marked on its page |
| GPT-5.6 Cyber | No effort sentence on its model page |
| GPT-5.5 Cyber | No effort sentence on its model page |
Reasoning modes on GPT-5.6 and GPT-6
| Pro mode | GPT-5.6 and GPT-6 families (Responses API): set reasoning.mode to "pro" |
|---|---|
| How pro mode bills | All the model work is billed at the selected model's standard token rates — more tokens, same rate |
| Reasoning tokens | Invisible in the API, but they occupy context and bill as output |
Reasoning effort is a published control on every current model — and this page used to say the opposite. Until the October re-check it described effort levels as published for exactly one model, GPT-5.3 Codex. The levels were on the other model pages all along, in a sentence shape this site's scraper couldn't parse; a Wayback snapshot of the GPT-5.5 page from before this site launched already lists them. The scraper now fails loudly on any effort sentence it can't read, instead of reporting the model as silent.
What the levels are
The full vocabulary across the roster is none, low, medium, high, xhigh and max, but no model accepts all of them in every family. GPT-5.6 models and GPT-6 Sol and Luna span none to max. GPT-6 Astra and GPT-6.1 Sol drop none — send low instead — and GPT-6.1 Sol's page names minimal as unsupported too. GPT-5.5 and the GPT-5.4 models stop at xhigh. The Pro models accept only medium, high and xhigh. The table above shows each model's exact set.
Defaults matter more than the levels
Most models mark medium as the default. The GPT-5.4 family marks none, which means an unconfigured request does no reasoning at all — a behavior change worth knowing before migrating to or from it. GPT-5.5 Pro defaults to high; GPT-6 Astra and GPT-5.3 Codex mark no default. In the Codex CLI, effort sits beside the model in /model, and the CLI's own default is medium on GPT-6.1 Sol.
What the dial costs
Higher effort means more reasoning tokens, which bill as output and occupy context even though the API never shows them. Pro mode on GPT-5.6 and GPT-6 does more model work again, billed at the model's standard rates. Codex's "Ultra" setting is a different thing: it uses maximum reasoning and may spawn subagents, so its cost depends on the tokens the whole task produces.
Verified 2026-10-28 against CodexHow facts module (src/data/facts/) — see /about/#accuracy.
Verified 2026-10-28 against https://developers.openai.com/api/docs/guides/reasoning.
Checked: https://developers.openai.com/api/docs/guides/reasoning · https://developers.openai.com/api/docs/guides/latest-model