--- title: "Model overthinking" description: "antiburn flags sessions that use an explicit reasoning or effort tier above its reviewed recommended cap." url: "https://antiburn.ai/docs/findings/model-overthinking/" --- ## What it detects Many agents let a model reason at different effort levels - low, medium, high, and above. antiburn reads the explicit tier a turn used (it never infers one from your prompt wording) and flags a session where any turn - main-loop or delegated to a subagent - ran at a tier above the reviewed cap: `xhigh`, `max`, or `ultra`. The cap is judged per model family, because the same label doesn't mean the same thing everywhere: Claude models recognize `low`, `medium`, and `high` as normal, with `xhigh`, `max`, and `ultra` above the cap. Codex (OpenAI) models recognize two lower tiers besides those, `none` and `minimal`, with the same three-tier ceiling flagged. This is a fixed rule, not a setting you adjust. ## Why it burns tokens Higher reasoning effort spends extra output tokens on every single reply, whether or not the task actually benefited from the extra thinking. Most tasks do fine on a lower setting, so the extra spend rarely buys better output. ## How to fix it Keep reasoning effort at `high` or below for most tasks in Claude Code, and reserve anything above that for the rare case that actually needs it. In Codex, the same ceiling applies, and you also have `none` and `minimal` available for genuinely simple work where even `low` is more than the task needs. ## When it shows "not assessed" If your agent doesn't expose an explicit reasoning-effort or thinking-mode setting at all, antiburn can't assess this - there's no tier field in the logs to read. For agents that do expose it, antiburn only judges turns that actually carry a tier value; a turn without one is skipped, not treated as "fine" by default. If no turn in the session reports a tier, it shows not assessed rather than clean - common on sessions with heavy delegated (subagent) work, since most harnesses write the effort field on main-loop turns but omit it on delegated ones. A session can also come back not assessed when a turn used a tier label no reviewed family recognizes, or partial log coverage means antiburn can't rule out an above-cap turn it never saw. See [hygiene score and findings](/docs/concepts/hygiene-score-and-findings/) for how this fits into your overall score.
Model overthinking
antiburn flags sessions that use an explicit reasoning or effort tier above its reviewed recommended cap.
What it detects
Many agents let a model reason at different effort levels - low, medium, high, and above. antiburn reads the explicit tier a turn used (it never infers one from your prompt wording) and flags a session where any turn - main-loop or delegated to a subagent - ran at a tier above the reviewed cap: xhigh, max, or ultra.
The cap is judged per model family, because the same label doesn’t mean the same thing everywhere: Claude models recognize low, medium, and high as normal, with xhigh, max, and ultra above the cap. Codex (OpenAI) models recognize two lower tiers besides those, none and minimal, with the same three-tier ceiling flagged. This is a fixed rule, not a setting you adjust.
Why it burns tokens
Higher reasoning effort spends extra output tokens on every single reply, whether or not the task actually benefited from the extra thinking. Most tasks do fine on a lower setting, so the extra spend rarely buys better output.
How to fix it
Keep reasoning effort at high or below for most tasks in Claude Code, and reserve anything above that for the rare case that actually needs it. In Codex, the same ceiling applies, and you also have none and minimal available for genuinely simple work where even low is more than the task needs.
When it shows “not assessed”
If your agent doesn’t expose an explicit reasoning-effort or thinking-mode setting at all, antiburn can’t assess this - there’s no tier field in the logs to read.
For agents that do expose it, antiburn only judges turns that actually carry a tier value; a turn without one is skipped, not treated as “fine” by default. If no turn in the session reports a tier, it shows not assessed rather than clean - common on sessions with heavy delegated (subagent) work, since most harnesses write the effort field on main-loop turns but omit it on delegated ones. A session can also come back not assessed when a turn used a tier label no reviewed family recognizes, or partial log coverage means antiburn can’t rule out an above-cap turn it never saw.
See hygiene score and findings for how this fits into your overall score.