OpenAI updated its Help Center usage documentation on October 31: from December, ChatGPT and Codex usage limits will be combined into a single allowance across the Plus, Pro and Business plans, metered in tokens. Some users had already spotted a similar notice on the checkout page, saying that from December a plan's usage would share one cap between ChatGPT and Codex.
The stated reason is that Work's agent capabilities are being folded into the main ChatGPT experience. OpenAI engineering lead Thibault Sottiaux had already hinted at the direction on a podcast in early October, saying the company plans to merge the Chat and Work interfaces into one. With the interface merging, the limits follow.
How many pools there are today
Right now, Work and Codex already share one agent allowance. Ordinary chat is not counted in it, and image generation, voice and file uploads each have their own limits and reset cycles.
According to media accounts of the Help Center page, the rules after December work like this:
- Usage is calculated on input and generated tokens, so more advanced models and more complex tasks use up more;
- Instant chats do not count toward the allowance, while the page does not detail how other ChatGPT modes are counted;
- The combined balance and reset time are shown on the Usage page under Settings;
- Once it runs out, users can wait for the reset, upgrade their plan, or buy credits for eligible features.
Some reports speculate that after the merge, image generation and long-running coding tasks may draw from the same allowance. Whether voice and file uploads are folded in too, OpenAI has not clearly said.
A back-of-the-envelope calculation
The token quotas under the new rules have not been published, so the only way to get an order of magnitude is to use current credit pricing. Prices compiled by third parties show GPT-6.1 Sol charging 50 credits per million input tokens and 250 credits per million output tokens.
Assume an ordinary Q&A exchange with 2,000 input tokens and 800 output tokens: that works out to roughly 0.3 credits. A Codex task that reads in 1 million tokens and writes out 200,000 comes to roughly 100 credits, about the same as 330 such Q&A exchanges. This is a rough estimate only; the allowance may not be converted at credit prices, and factors such as caching are not included.
The gap points to one thing: heavy coders and chat-only users will experience this merge very differently. A chat-only user who often picks a high-end model like Sol may, for the first time, see their chats draining a balance; a user who runs dozens of long Codex tasks a week will find chat starting to compete with those tasks for the same budget. The model tiers are also far apart: in estimates likewise from third parties, Plus gets 15 to 160 Sol messages every five hours, versus 350 to 3,000 for Luna.
OpenAI's position is that most users will not hit limits more often than they do now. That is an expectation, not something the Help Center page frames as a guarantee.
What is still unspecified
The update leaves quite a few gaps:
- Which day in December the switch happens, and whether it is done all at once or in batches;
- How many tokens each plan includes;
- How the existing five-hour windows and weekly caps apply to the new allowance; a third-party compiled page shows that Pro currently has no five-hour limit;
- Whether Free, Go, Enterprise and Edu will fall under these rules.
The news is also part of the Codex team's 28-day plan. On October 4, Sottiaux committed that every day for 28 days the team would either ship an improvement most users could use or run a full reset; as the stretch drew to a close, reports mentioned the team issuing another reset. Short-term resets and doublings are perks, but it is the post-December quota numbers that determine what a subscription is actually worth. Until OpenAI publishes them, there is no way to tell whether the merge adds to or subtracts from what each tier gets.
Sources: OpenAI Help Center usage documentation, Windows Report, CocoLoop, Startup Fortune; credit prices and message counts are figures from current third-party compiled pages, and the token conversions are the editors' rough estimates.