When Usage Limits Become a Product Feature: Codex, Tibo, and Claude
Recently, the Codex conversation has been less about a new model and more about one recurring action: resetting usage limits.
Sometimes the reset follows an incident. Sometimes it follows a cache or accounting problem, a product launch, or a user milestone. The five-hour limit has also disappeared from some interfaces, leaving only the weekly balance. Meanwhile, paid users have repeatedly seen their weekly allowance restored. Claude reset five-hour and weekly limits for all users on the same day GPT-5.6 launched.
Then Codex leader Tibo Sottiaux replied to Claude's announcement with three words: “I smell fear.”
It was an excellent screenshot. But beneath the joke is a more important shift: in the agent era, quotas, resets, and rate limits have moved from backend infrastructure into the product experience itself.
The five-hour limit disappeared—but limits did not
In July, OpenAI support explained in its developer forum that the five-hour limit display had been temporarily removed as part of incident response, while weekly reset information would remain available. Read the OpenAI discussion.
That distinction matters. Seeing no five-hour limit should not be read as a permanent cancellation of the rolling window. The safer interpretation is that one dimension of the quota system was temporarily hidden or paused while the backend policy remained in place.
For users, however, the practical difference is small. A previously clear two-layer system—five-hour window plus weekly quota—becomes a single weekly number. That creates basic questions: how much can I use now, is the recovery rolling or weekly, is 100% a one-time compensation, and which surface is authoritative when the CLI, desktop app, and web dashboard disagree?
Those are workflow questions, not documentation trivia.
Why “reset” increasingly looks like a product button
Rate limiting used to be quiet infrastructure: reach the ceiling, wait, or upgrade. Codex has developed a different rhythm: an incident, a reset; a new model, a reset; a user milestone, a reset; reports of unusually fast depletion, another reset.
Some of these actions are responsible compensation. OpenAI has said that reports of Codex limits depleting too quickly were partly related to abuse and fraud-prevention systems incorrectly rate-limiting accounts. OpenAI Status.
The problem begins when the same compensation mechanism appears so often that it becomes part of how users understand the product. The community has started tracking Tibo's public resets across incidents, cache issues, GPT-5.6 Sol, and user-growth milestones. Public timeline.
Users do not dislike free quota. They dislike not knowing what the rule will be next week.
“I smell fear” and the retention battle
On July 9, ClaudeDevs announced that it had reset five-hour and weekly limits for everyone, at almost exactly the moment GPT-5.6 was being released. Tibo's reply was interpreted as: Anthropic saw competitive pressure and handed users a fresh refill. ITmedia's report.
The reply does not prove Anthropic's motive. Capacity changes, incident compensation, and normal operating policy are all plausible explanations. It spread because it compressed an unproven causal story into three dramatic words.
This is no longer a static model comparison. It is a dynamic retention contest: model launches attract users, while quota resets give them a reason to stay and run one more task.
The real unit of competition: each accepted change
For developers working in real repositories, the experience is shaped by four variables: whether a task finishes before the quota runs out; whether context can be recovered after interruption; whether the next available window is predictable; and whether quota consumption matches actual output.
That is why “cheap” cannot be measured only by token price, and “generous” cannot be measured by whether a reset happened on a particular day. A more useful metric is cost per accepted code change: money, interaction rounds, tool calls, and human rework.
Track the starting balance, model, effort level, tool-call count, and end reason for each task. Set budgets for agent workflows. Use tests, linting, and type checks as evidence instead of forcing endless self-verification. Give long-running tasks recoverable checkpoints.
The reset button is a benefit—and a liability
Codex hiding the five-hour limit, OpenAI repeatedly restoring weekly quotas, and Tibo joking about Claude are parts of the same contest: who can turn expensive agent compute into a daily productivity product people are willing to keep paying for.
A reset can repair an incident or turn a model launch into a real trial. But it also creates product debt. Once users expect someone to press the button whenever the system becomes uncertain, stable rules become harder to trust.
We do not need to decide whether Tibo was right about fear. The more important point is this: the next stage of AI coding competition will be fought not only over model capability, but over quota transparency, recovery mechanisms, and whether users can safely hand ongoing work to an agent.
Free quota makes users happy for a moment. Predictable quota is what makes them stay.