---
title: "Fast mode overuse"
description: "antiburn flags sessions where subagents ran on a fast speed tier instead of standard."
url: "https://antiburn.ai/docs/findings/fast-mode-overuse/"
---
## What it detects
Some agents let you run a model in a faster, higher-throughput speed tier alongside its normal ("standard") tier. antiburn watches for that fast tier specifically on delegated work - turns a subagent ran, not the main thread - and flags a session once delegated fast-tier turns reach antiburn's reviewed threshold, which is a single delegated fast turn.
The rule only looks at delegated work today, because antiburn doesn't yet have a reliable signal for a session's standing default speed on the main thread. A main-loop turn running on fast mode, by itself, is not a finding.
## Why it burns quota
Fast mode trades a materially higher per-token rate for lower latency. That trade can be worth it for a burst of work where speed matters, but paying the fast-mode premium as the default for routine subagent work adds up fast for a small speed gain most delegated tasks don't need.
## How to fix it
Use standard speed by default, and reserve fast mode for the rare subagent task where the latency actually matters. In both Claude Code and Codex, subagent runs recognize the same two speed labels - `fast` and `standard` - so the fix is the same in either: check whatever setting controls the subagent's speed tier and default it to standard.
## When it shows "not assessed"
If your agent doesn't expose a fast/standard speed distinction at all, antiburn can't assess this - today that's every supported agent except Claude Code and Codex.
For Claude Code and Codex sessions, a session can still come back not assessed rather than clean in a few ways: the session used a speed label that doesn't normalize to `fast` or `standard`, so antiburn can't classify it; no turn in the session reported a speed value at all, since antiburn only judges turns that actually carry one - a turn without the field isn't assumed to be "standard"; or the source only carries a coarser "service tier" marker antiburn's evidence contract can't yet read into a fast/standard verdict. Partial model evidence blocks a clean result the same way, even with no finding observed.
See [hygiene score and findings](/docs/concepts/hygiene-score-and-findings/) for how this fits into your overall score.
Fast mode overuse
antiburn flags sessions where subagents ran on a fast speed tier instead of standard.
What it detects
Some agents let you run a model in a faster, higher-throughput speed tier alongside its normal (“standard”) tier. antiburn watches for that fast tier specifically on delegated work - turns a subagent ran, not the main thread - and flags a session once delegated fast-tier turns reach antiburn’s reviewed threshold, which is a single delegated fast turn.
The rule only looks at delegated work today, because antiburn doesn’t yet have a reliable signal for a session’s standing default speed on the main thread. A main-loop turn running on fast mode, by itself, is not a finding.
Why it burns quota
Fast mode trades a materially higher per-token rate for lower latency. That trade can be worth it for a burst of work where speed matters, but paying the fast-mode premium as the default for routine subagent work adds up fast for a small speed gain most delegated tasks don’t need.
How to fix it
Use standard speed by default, and reserve fast mode for the rare subagent task where the latency actually matters. In both Claude Code and Codex, subagent runs recognize the same two speed labels - fast and standard - so the fix is the same in either: check whatever setting controls the subagent’s speed tier and default it to standard.
When it shows “not assessed”
If your agent doesn’t expose a fast/standard speed distinction at all, antiburn can’t assess this - today that’s every supported agent except Claude Code and Codex.
For Claude Code and Codex sessions, a session can still come back not assessed rather than clean in a few ways: the session used a speed label that doesn’t normalize to fast or standard, so antiburn can’t classify it; no turn in the session reported a speed value at all, since antiburn only judges turns that actually carry one - a turn without the field isn’t assumed to be “standard”; or the source only carries a coarser “service tier” marker antiburn’s evidence contract can’t yet read into a fast/standard verdict. Partial model evidence blocks a clean result the same way, even with no finding observed.
See hygiene score and findings for how this fits into your overall score.