Reusing context and other changes to slow usage-limit consumption
Thibault Sottiaux, who leads OpenAI's AI coding tool Codex, posted on X on Oct. 16 (Oct. 17 Japan time) that Codex and ChatGPT Work had received improvements that increase the amount of work that can be done under the same usage limits. Context is now reused between subagents, the helper AIs that split up a task, and the cache (a mechanism that reuses input that has already been read) is kept longer, while the routing that decides which server or model handles a request was revised. He said the cache hit rate and the effective usage limits had improved, but gave no figures on how much usage would fall.
In Codex, a parent AI often splits a single task among multiple subagents. If subagents inherit the parent's entire conversation history, the same context is read again and again, and the usage limit is used up faster. Input that hits the cache is counted at a lower rate than usual, so a higher hit rate allows more work under the same limit.
A pledge of an improvement or a reset every day for 28 days
The improvement was released as day 12 of "28 Days of Shipping," which OpenAI ran from Oct. 5 to Nov. 1. Sottiaux had pledged to ship one visible improvement every day during the period or else reset usage limits for all users. On the same day, ChatGPT's "GPT-6 Astra Pro" gained the ability to run multiple agents in parallel, and the always-on agent "dots" could be placed on the home screen. On the following day, Oct. 17, OpenAI reset usage limits for all users instead of shipping an improvement.
According to secondary sources, Codex and ChatGPT Work reached a combined 40 million users as of Oct. 7. On Aug. 29, OpenAI also fixed eight bugs that consumed extra usage and said weekly usage limits would last 10% to 50% longer depending on how they are used. At that time it also fixed a bug in which a small model called subagents running higher-tier models without being instructed to.
| Date | Change |
|---|---|
| Oct. 5 | Day 1. Standard responses made about 50% faster |
| Oct. 16 | Context reuse between subagents, longer cache retention, improved routing |
| Oct. 17 | Usage limits reset for all users |
| Oct. 20 | API prompt cache kept for up to 12 hours |
| Oct. 23 | Cache kept even when reasoning depth is changed. Codex CLI updated |
| Oct. 29 | Usage limits doubled |
No size of the gain given, and complaints about subagent consumption
OpenAI has not said how much the improvements reduce usage. On Oct. 13, a report that subagents consume a lot was filed in Codex's public developer repository. In the case of a user who connected a non-OpenAI model, three subagents that inherited the parent's full history used about 11.06 million tokens in a day, nearly half of the roughly 24 million tokens used in total, according to the report. The same reporter also wrote that the cache hit rate was 96.9% with OpenAI's models, noting that the comparison was not made under matching conditions.
Codex CLI 0.163.0, released the day before on Oct. 15, added a setting to choose the default context passed to subagents and the ability to run read-only tools in parallel. Version 0.164.0 on Oct. 23 changed how the working environment is handed over to subagents and how prompt cache keys are handled.
On Oct. 20, OpenAI acknowledged that about 40,000 accounts had been given usage limits different from their chosen plans over the weekend, and apologized. Starting in December, it plans to combine usage limits for ChatGPT and Codex into a single allowance. Anthropic also removed the five-hour limits on its top-tier plans on Oct. 30, making usage limits a focal point of competition between the two companies.
