OpenAI reportedly confirmed a bug that applied wrong usage limits to about 40,000 Codex accounts

What happened

According to a report on dev.to, developers began noticing over the weekend of October 17–18, 2026 that their Codex allowance was being consumed far faster than comparable work had cost a week earlier. On October 20, Tibo Sottiaux — described in the report as OpenAI's lead for Codex and ChatGPT — responded publicly on X. He said that, prompted by the reports, OpenAI had identified roughly 40,000 accounts where the limits actually being enforced did not match the plan each user had selected, that the problem had been corrected, that affected users were being contacted individually, and that credits would be added to those accounts.

The dev.to author adds a caveat worth keeping in mind: when the post was re-checked on October 28 against OpenAI's Help Center, the October 20 statement could not be verified on an official OpenAI page, and the exact wording of the post was not re-confirmed. The account should therefore be read as reported rather than as documented in OpenAI's official support pages.

What OpenAI has not said

The gaps matter as much as the admission, according to dev.to. OpenAI has not named the plans involved, has not said how large the credits will be or when they will arrive, and has not published any way for a user to check whether their account is among the affected 40,000 — there is no claim to file, since contact is meant to come from OpenAI's side. The company has also not linked the bug to every complaint about fast-draining quota, including separate reports about GPT-6.1 Sol consumption, so some users hitting a wall are presumably seeing genuine limits.

How Codex limits work

Codex has no fixed messages-per-day number. OpenAI's Help Center, as cited by dev.to, describes published figures as estimates that vary with the model, the size and complexity of the task, how much context the agent carries, reasoning effort, tool use, and whether work runs locally or in the cloud. Two structural points shape the experience: Codex and ChatGPT Work draw from a single shared allowance, and local messages and cloud tasks consume the same pool, with cloud tasks typically costing more.

Metering is split across two windows. On Plus and Standard Business plans, usage is measured in a rolling five-hour window, with a weekly cap on top. Current Pro plans have no five-hour limit at all — only the weekly allowance applies. Burn rate also depends on settings: per the Help Center, Fast mode consumes included usage at 2.5 times the Standard rate and Ultrafast at 8 times, while image generation uses allowance roughly three to five times faster than a similar text-only exchange. Large AGENTS.md files and enabled MCP servers add context cost to every request.

Telling the bug from a real limit

The dev.to guide lists signs that point to a metering error rather than genuine consumption. A meter that leaps instead of climbing steadily is one: a public GitHub report from October 16 described a Pro weekly meter moving from 19% to 75% in about fifteen minutes of ordinary work. Another is usage attributed to a model the user never selected — one Pro user found roughly 80% of a week's usage charged to GPT-6 Astra in the analytics view while every individual chat showed GPT-6.1 Sol. Contradictory meters are a third: users reported Codex code reviews refused for exceeding limits while the usage page showed half the weekly allowance still available. A drain that began during the October 17–18 weekend with no change in workflow also fits the bug's timeline.

By contrast, a real limit usually has an explanation in your settings: Fast or Ultrafast modes, maximum reasoning, image generation, or long high-context sessions can exhaust an allowance quickly at the multipliers above.

Checking actual usage

Two official views should agree, per the guide: the usage dashboard at chatgpt.com/codex/settings/usage (or ChatGPT Settings → Usage), and the /status command inside a Codex CLI session, which shows remaining limits for the active session. The desktop app's Settings → Usage & billing adds a per-model breakdown, which is what exposed the misattributed usage in one of the reports. The Help Center also states that Support cannot reset limits. OpenAI's own advice is to check the dashboard every week or two to establish a personal baseline, which makes an abnormal drain obvious when it happens.

Why it matters

Paying developers were locked out of a tool they had budgeted for, and the remedy on offer — individual outreach plus unspecified credits — leaves them unable to verify whether they were affected at all. For agentic coding tools, metering is effectively the pricing interface: when the meter is wrong, users can no longer trust either their bills or their own efficiency judgments. The episode also illustrates a verification problem, since the confirmation reportedly lives in a social media post rather than the Help Center, making it hard to check after the fact. As coding agents become metered utilities, transparent and independently checkable usage accounting will be table stakes.