Codex usage limit: 5-hour cap gone, Anthropic responds

On July 12, OpenAI removed the Codex usage limit: the 5-hour cap that set the working pace for everyone using the agent on the Plus, Pro and Business plans. No blog post, no keynote. A member of the technical staff announced it on X and, in less than an hour, the usage counters were reset.

Anthropic answered within hours. In the early hours of July 13, it extended through July 19 the promotion that raises Claude Code's weekly limit by 50%, and stretched the Fable 5 access included in paid plans along with it.

For the first time, the two labs are competing over who lets you use the agent more, not over who looks better on the benchmark. In this post: what actually changed, what is still capped (spoiler: the Codex weekly limit isn't dead) and how to redo the math if you pay for both subscriptions.

TL;DR

What changed: OpenAI temporarily removed the 5-hour cap on Codex and ChatGPT Work; Anthropic extended the 50% increase to Claude Code's weekly limit through 07/19.
Who benefits: Plus, Pro and Business on the OpenAI side; Pro, Max, Team and seat-based Enterprise on the Anthropic side.
Cost/Access: zero. Everything is applied automatically, no plan upgrade needed.
Useful link: discussion in the official Codex repository confirming that the 5-hour meter disappeared from the app.

What OpenAI changed in the Codex usage limit

On Saturday, July 12, a member of OpenAI's technical staff announced on X the temporary removal of the 5-hour limit for all Plus, Business and Pro users, on both Codex and ChatGPT Work. The counter reset followed in less than an hour.

The context explains the rush. On the 11th, GPT-5.6 Sol, the heaviest reasoning tier of the new flagship, became broadly available, and the high compute mode burned through the usage budget far faster than users expected. Mass complaints, quota evaporating, and OpenAI opened the tap. Along with removing the cap, the company promised efficiency optimizations for Sol: the same work should start consuming less quota.

The detail that matters: it's temporary. OpenAI didn't announce a date for the limit to come back, but it didn't promise the limit is gone for good either. Anyone reorganizing their workflow around this is building on a promotion, not on a contract.

What is still limited (read before celebrating)

The 5-hour cap was the short-window limit. There is a second, separate system: the weekly limit, and that one is still standing. Resetting one doesn't restore the other.

And there's the structural catch that removing the cap doesn't fix: ChatGPT Work and Codex share the same usage pool. A long session by someone on your team in Work draws from the same budget as your coding agents. There's a documented case of a Plus user watching their quota go from 70% to 100% consumed in about six minutes of ordinary use. Not because of a bug, but because of the billing architecture.

Translation: OpenAI opened up the short window and kept the weekly ceiling. You can run faster, but the marathon is still the same distance.

Anthropic's answer: hours, not weeks

Anthropic had already been raising Claude Code limits at the cadence of a company under pressure: it doubled the 5-hour limits in early May and, a week later, raised the weekly limits by 50%. That's 2x on the short window and 1.5x on the weekly one, with no price increase. The promotion had an expiration date: July 13.

Then OpenAI dropped its announcement on the 12th.

In the early hours of July 13, hours later, Anthropic extended the promotion through July 19 at 11:59 PM PT. It applies to Pro, Max, Team and seat-based Enterprise (legacy). The Free plan and usage-based Enterprise are left out. No action required: the higher limit is applied to the account automatically.

In the same move, the Fable 5 access included in paid plans was extended for the second time, also through July 19. The previous deadline, July 12, had already passed when the announcement went out. Days earlier, on July 10, the company had already reset the 5-hour and weekly counters for all users after an unstable week.

An extension announced after its own deadline, hours after the competitor's announcement, is not a calendar coincidence. It's a retention war.

▪ Clã Beer and Code

Do not just follow the news — master it. Hands-on AI Engineering, live, every week, in the largest community in Brazil.

Join the Clã

Which subscription is worth it right now?

The honest answer: this week is the best window of the year to measure, not to decide.

If you pay for OpenAI's Plus or Pro, Codex is running without the short cap. You can run those long agentic workflows that used to die halfway through and see how much of the weekly limit they really eat. If you pay for Anthropic's Pro or Max, you have 50% more weekly ceiling on Claude Code until the 19th. You can find out whether your real usage fits in the lower plan or demands the higher one.

Write down your consumption on both sides during the week. When the limits go back to normal (and they will), you decide with numbers, not with a Reddit anecdote. The real Claude Code pricing table for 2026 (Pro, Max, API) is broken down in Did Claude Code get 5x more expensive?, and the technical comparison between the two agents on a real repo is in Which AI to use for coding in 2026.

A warning from someone who lives off the SERP: if you search for "codex usage limit" today, a good share of the results still explain the 5-hour cap as if it were in effect. There's an article from September 2025 at the top of Google describing in detail a limit that just stopped existing. News about AI limits ages in days, not months. Always check the date.

Limitations and things to watch

Everything here is temporary. Anthropic's promotion has a date and time to die (07/19, 11:59 PM PT). OpenAI's removal has no date either to end or to become permanent, which is worse for planning, not better.
The weekly ceiling exists on both sides. Don't treat the week as unlimited usage: an agent running in an unattended loop eats weekly quota just the same.
Don't couple your pipeline to a promotional limit. If you have automation with Codex or Claude Code in CI, treat rate limits as a normal scenario: retry with backoff, fallback and alerting. A promotion is not an SLA.
OpenAI's shared pool still applies. A team using ChatGPT Work and Codex on the same account keeps splitting the budget. Removing the cap didn't separate the pockets.
A usage limit is not a context limit. They are different axes, and it's easy to confuse the two. A few days after touching usage, OpenAI also cut Codex's context from 372k to 272k tokens: another cap, another kind of pain. If your agent keeps forgetting the beginning of the task, it's not the quota. That's what we break down in Codex with reduced context.

Quick FAQ

Is Codex unlimited now?

No. The 5-hour cap was temporarily removed, but the weekly limit is still active and it's a separate system. On top of that, Codex and ChatGPT Work share the same usage pool: consumption on one side is deducted from the other.

How long does the 50% increase to the Claude Code limit last?

Until July 19, 2026, at 11:59 PM Pacific Time. It applies to Pro, Max, Team and seat-based Enterprise. It's automatic, no configuration needed.

Do I need to do anything to activate the changes?

No, on either side. OpenAI reset the counters at the announcement; Anthropic applies the extended limit directly to eligible accounts.

Is it worth subscribing to both?

It depends on your usage profile, and this week is the ideal moment to find out, because both have the bar set high. The criteria for choosing by scenario (large repo, terminal, cost) are in our Claude Code vs Codex comparison.

The game has changed: now the fight is over retention

For two years, the war between the labs was about benchmarks. Now it's about how much you can use before hitting the ceiling, because the coding agent became the product that pays the bills, and a dev who hits a rate limit three times in the same week cancels and subscribes to the competitor.

The next chapter is predictable: either OpenAI makes the removal permanent and forces Anthropic to rethink the weekly limit, or the cap comes back and this week becomes a footnote. This is a living post: if either of them responds again, we'll update it here.

In the meantime, the rational play is to use the window to measure your own consumption and decide on a subscription with numbers in hand. It's exactly the kind of test we run every week, live and on production code, at Clã Beer and Code, the largest AI engineering community in Brazil. The limits war changes every week; what doesn't change is that those who measure decide better than those who cheer.