The edit appeared in OpenAI’s Help Center over the weekend, under a heading that reads like a line from a utility bill: “Understanding ChatGPT and Codex usage limits.” By Monday morning, developers who pay for both products had read it more than once.
Starting in December, OpenAI will merge the allowances behind ChatGPT and Codex for Plus, Pro and Business subscribers into a single pool, according to the updated article. Usage will be counted in tokens rather than in the approximate message counts the plans describe today. A long stretch of agentic work in ChatGPT will leave less room for code review in Codex. The reverse is also true.
The exemption is the detail most subscribers will notice first. Chats handled by OpenAI’s Instant models will not draw on the shared allowance, the company said. Everything else is weighted by model and task: stronger models and more tangled requests consume the pool faster, and the amount a single request uses depends on both.
What OpenAI has not published is nearly as interesting. There is no start date inside December. There are no new allowance sizes for any plan. And the article does not say how the five-hour and weekly windows that govern today’s plans will carry over to the combined balance. For teams that plan capacity the way they plan cloud storage, those are the numbers that decide whether a subscription still fits.
The mechanical change is modest and the commercial one is not. Cloud-based Codex already draws on a pool shared with ChatGPT Work, and OpenAI’s Codex pricing page says Work usage inside ChatGPT “uses the same pricing, credits, and usage limits as Codex.” December widens the pool to cover the whole of ChatGPT for the three paid consumer and small-business tiers. Tokens replace message estimates as the unit.
The stated reason is product structure. OpenAI has been folding the agentic capabilities of Work into the main ChatGPT experience, and it has said the separate Chat and Work tabs will be merged, first for a small test group. Once a single interface can run a two-hour research task and answer a question about a spreadsheet, the billing boundary between “chat” and “agent” stops describing anything real. One meter is the accounting version of that merger.
The timing follows a round of generosity. On October 29, OpenAI doubled the Work and Codex allowances for its Pro 100 and Pro 500 tiers and canceled a planned reduction for Pro 200 subscribers, a change confirmed in its usage documentation. Subscribers who had already received the higher allowance kept it. The December pooling is a separate change, but the two arrive close enough that users will read them together.
Analysts said the shift is less about extracting more revenue per seat than about making variable costs legible. Inference is billed by the token to OpenAI, and a flat monthly fee against wildly variable agent usage is a margin problem the company has been managing with unpublished limits for a year. A shared, token-denominated pool converts that problem into a number customers can see, and into a reason to buy credits when the number runs out.
For engineering organizations, the practical question is forecasting. A team that runs Codex for pull-request review in the morning and ChatGPT for incident analysis in the afternoon now draws from one balance with an unfamiliar burn rate. The old message counts were crude, but they were stable enough to model. OpenAI’s guidance that most users will not wait for resets more often than they do now is a projection about typical behavior, not a floor on what any given account receives.
Procurement teams have a second question. The change covers Plus, Pro and Business. OpenAI has not announced the same treatment for Free, Go, Enterprise, Edu or Healthcare accounts, which leaves larger buyers outside the pool and, for now, outside the confusion. Several of those contracts are already priced per token, so the December change touches the segment of the market that buys seats rather than compute.
Competitors are watching the same seam. Anthropic sells Claude through subscription tiers with their own message allowances and has been steadily tightening the language around what counts against them. Google bundles Gemini into Workspace seats at a flat price. Whichever vendor first publishes an allowance big enough to be planned against, rather than large enough to be marketed, wins the spreadsheet.
Inside OpenAI, the change also simplifies a support burden. Users who hit a wall currently have three options: wait for the reset, upgrade, or buy credits to keep using eligible features. Each of those paths needed its own explanation depending on which product and which window applied. A single balance in Settings, under Usage, with one reset time, is a smaller help article and a cleaner upsell.
None of this settles whether the allowances themselves are adequate. OpenAI has not said whether Plus subscribers will get the same effective capacity in December as they had in November, and it has declined to describe how the pool is split across models. Subscribers who use ChatGPT mainly for Instant conversations lose nothing. Subscribers who discovered agentic work this year are the ones whose monthly ceiling is about to become a number they can watch fall.
The Help Center article runs to a few hundred words and ends with a note that advanced features are weighted differently. That sentence will be quoted back to OpenAI’s support team many times in December, by people asking what “differently” means for the twenty dollars or two hundred dollars they pay each month.
