Codex Switched to Tokens: Why Limits Run Out Within Hours

⚡ The gist of the dispute

A thread with 1,023 comments has opened in the OpenAI Community: Codex has been switched from message counting to token counting, and the promo with doubled limits has ended. Users report that quotas now run out sooner and compare who spends them on what. The post author jochenschultz lists consumption factors: large prompts, big repositories, long context, bulky AGENTS.md files, MCP servers, Fast mode, image generation, and model choice.

🗣 What participants are saying

A user under the nickname John_Diesiel sees a recurring scenario: "First Antigravity usage reduction. Now Codex". In his view, tech giants first attract users, let them build their workflow around the tool, and then cut access and call it progress. His conclusion is harsh: "From the outside, it looks greedy", and from the user's side — "unstable, manipulative, and arrogant".

soakes calls the phrase "too restrictive" an understatement: according to him, it's "a bit of an understatment", and the system is "literly unusable for business and personal accounts". He speaks about both business and personal accounts without providing a separate measurement.

qmqz describes a specific scenario: on the Plus plan he "managed to wipe out my weekly usage in a few hours" while working with 5.4 medium in IntelliJ IDEA. He doesn't dispute the quality of the results — "the work it accomplishes is great, but it lasts very briefly" — but can't understand "how this can possibly be worth paying $20 a month".

rob.graham.au is close to canceling his subscription: tasks that went through fine a few days ago are now "chewing not only 5hr rate but weekly rate". He doesn't name the model or environment, but records a shift on both limits — the five-hour and the weekly one.

dunderfoot compares it with Claude: there, in his experience, it's "way more expensive and not as good". His remark — "Now codex is like claude" — is also about degradation, but through the price and quality of the alternative rather than through token mechanics.

⚖️ Where they agree, where they diverge

The common denominator in the quoted comments is one: after the switch to tokens and the end of the promo, consumption is perceived as faster. The complaints converge on the weekly quota and on the "5.4 medium + IntelliJ" combo for qmqz, on the five-hour limit for rob.graham.au, and on unsuitability for work for soakes. At the same time, these are five remarks out of 1,023 comments — not the thread's consensus, but a set of positions.

The real conflict isn't in token arithmetic but in how to evaluate the deal. John_Diesiel talks about deliberate restriction of access, qmqz and rob.graham.au about the practical failure of familiar tasks to pay off, soakes about unsuitability for business. dunderfoot doesn't defend the new mode but poses the question differently: the alternative is more expensive and weaker. There is no shared answer in the comments as to where the honest boundary of the limit lies.

🧪 What to check yourself

• Run the same task on 5.4 medium in IntelliJ and in the CLI, recording the consumption of the five-hour and weekly quota before and after.
• Trim AGENTS.md, disable unused MCP servers, and compare consumption on the same repository.
• Check how much Fast mode and image generation eat up, if you use them.
• Calculate the cost of one completed task on the $20 Plus plan and compare it with the alternative mentioned by dunderfoot.
• For a work scenario, check in advance whether it fits within the weekly limit on a test set of tasks, not during the promo period.

🔗 Sources

OpenAI Community — Codex Rate Limit Discussion Thread

Disclaimer: this is a summary of a discussion, not facts verified by the editorial team. The opinions belong to the participants, the numbers to those who cited them.