✗ Codex usage limit draining 10-20x faster than normal

There is no error string for this one — the symptom is your usage meter. Since mid-June 2026, Codex users across plans report limits burning far faster than the same work cost before. Issue #14593 (“Burning tokens very fast”) is the repo’s most-commented open issue at 627 comments, and issue #28879 (550 reactions) quantifies the jump from session logs.

Symptoms

• A Plus 5-hour budget gone in 2-3 prompts that previously allowed 20+, per #28879. Its logs: a ~57K-token prompt cost about 1% of the budget on June 12; ~20K-token prompts cost 10-27% each on June 18 — “the limit-% consumed per token has increased roughly 10–20×.”
• One Plus account went 70% to 100% in about six minutes on July 2.
• A $200 Pro weekly limit exhausted in roughly two days on 5.5, where 5.4 rarely came close — our models reference covers what changed between them.
• Idle drain: one report on #28879 describes sending a single greeting, walking away, and returning to find 20% of the 5-hour limit gone.

What causes it?

No official root cause. The #28879 reporter correlates onset with June 16, rules out local causes (reasoning effort, context bloat, config), and asks whether per-token limit weighting changed server-side. One contributor’s analysis on #14593 argues newer builds batch fewer tool calls per model cycle (roughly 1 versus 1.4-1.8 before), inflating input tokens per tool action about 3.5x — community analysis, not confirmed by OpenAI. OpenAI has assigned #28879 to an engineer and labeled both threads rate-limit bugs, but we found no public maintainer response on either.

How do you fix it?

Community-reported mitigations, nothing official:

1. Update Codex and restart. Reporters on #14593 describe the drain quietly stabilizing after updating (one dated June 12), and a #28879 commenter reported limits back to normal on July 12 without pointing to a cause. No release note claims a fix.
2. Disable memory and prompt-recommendation features. The one traced idle-drain case on #14593 — quota falling from 99% to 0% with no interaction — stopped after turning those off.
3. Trim per-turn context. If the batching analysis holds, every action carries more input tokens. A lean AGENTS.md and earlier compaction reduce exposure.
4. Confirm you’re on the right page. If the dashboard shows quota remaining while Codex blocks you entirely, that’s a different bug.
5. Add your logs to the tracking issues. Before/after token_count and rate_limits lines are what made #28879 credible.

Which versions are affected?

No clean version range exists. Reports span the VS Code extension, desktop app, and CLI, going back to March 2026; the quantified 10-20x jump clusters after June 16, 2026. Our usage-limits reference tracks the per-plan numbers, and we’ll update this page as the threads move.

Facts on this page verified July 15, 2026.