ChatGPT and Codex Move to Token Allowances in December

Starting in December, ChatGPT Plus, Pro and Business subscribers will draw ChatGPT and Codex usage from a single shared allowance, measured in tokens. That is the substance of a short OpenAI help article, reported live by October 31, 2026, that gives no day in December, no token figures and no per-plan amounts. It does say which usage counts, which does not, and what happens when you run out.

The change matters because it alters the unit. Today a paid plan is described in messages: so many with one model, so many with another, over a rolling window. From December the unit is the token, which scales with the length of what you send and receive, the model you pick and how hard it thinks. A long agentic Codex session and a short chat will no longer count the same, and a team that budgets seats by message counts will need a new mental model.

This post is deliberately literal. Every claim about the December change is a sentence from the article; everything about plan prices and consumption rates comes from OpenAI's pricing page and its Pro tiers article, read on November 5, 2026. Where OpenAI has not said something, we say that it has not, rather than fill the gap.

The source is a help-centre article titled "Understanding ChatGPT and Codex usage limits". It is three paragraphs long, and because every later claim rests on it, the content is worth setting out sentence by sentence rather than summarised. The article says that "Starting in December, Plus, Pro, and Business plans' usage allowance will be shared across ChatGPT and Codex as we bring the agentic capabilities of Work into the main ChatGPT experience."

It then explains the unit. Usage for ChatGPT and Codex will be measured in tokens, which it defines as the units of information a model processes and generates. How much of an allowance a task consumes will depend on the model chosen and the task asked. "More advanced models and complex tasks generally use more of your allowance. Instant chat will not draw from your usage."

The third paragraph is about expectations and mechanics. OpenAI says it is preparing carefully to prevent disruption, that "We expect most users won't need to wait for their usage to reset more often than they do now", that combined remaining usage and reset times will be visible under Settings, Usage, and that a user who reaches the limit can wait for it to reset, upgrade, or buy credits to continue using eligible features. That is the entire public record of the change as of November 5, 2026.

What is absent is as important as what is present. No day in December. No token amounts for any plan. No statement about Free, Go, Enterprise or Edu. No price change. No definition of which features are "eligible" for credits beyond what existing help pages already say. The dateline is also soft: the article page shows no publication date, and the October 31 date comes from when it was first seen live, not from the page itself.

The article names three plans and two products. The products are ChatGPT and Codex, with the agentic capabilities of ChatGPT Work, OpenAI's agent product, folding into the main ChatGPT experience as the reason the pools merge. The plans are Plus, Pro and Business. Below are those plans as OpenAI's pricing page and Pro tiers article describe them on November 5, 2026, with the plans the article does not mention listed for completeness.

Two observations from the table. First, the three named plans are exactly the self-serve paid tiers that include both products in volume, so the change reaches the people most likely to run an agent for an hour. Second, the silence on Enterprise and Edu is consistent with how OpenAI already prices them: its Codex pricing page says flexible-pricing Enterprise and Edu customers have no fixed rate limits and scale with credits, while those without flexible pricing carry the same per-seat limits as Plus. Whether that second group moves to the shared token pool is not stated. Treat it as open.

A message is a count. A token is a measure of work. The difference is why the change deserves a post rather than a footnote. Under a message limit, a one-line question and a request to refactor a repository each cost one message, and the allowance ran out at the same count regardless. Under a token allowance, the refactor costs what it costs: every file the agent reads, every tool result it processes and every line it writes is metered, and the model's own reasoning adds to it.

OpenAI has already described what drives that cost, in its Pro tiers article rather than the December one. Asked why similar tasks use different amounts of included allowance, the article answers that usage depends on the model, the amount of context processed, reasoning, tool use, information retrieval, caching and speed mode. That list is the best public description of what a token-metered allowance will respond to, and each item maps to a habit a user controls.

OpenAI has not published token allowances, but it has published the rates at which some choices consume whatever the allowance turns out to be. They live in the Pro tiers article and apply today to Pro plans; they are the clearest signal of how the December pool will behave, because they are already expressed in allowance terms rather than messages.

The example that follows from the table: a Pro 500 user who runs an agentic session in Ultrafast consumes included allowance eight times faster than the same session in Standard. Whether that is worth it depends on the hourly value of the person waiting, which is the calculation our analysis of what $200 a month of AI buys walks through. The failure case is leaving Fast mode switched on by habit for work that nobody is watching, at two and a half times the burn.

The temptation with a change like this is to back out a number: take today's message ranges, multiply by a typical message length, and announce a token allowance. We are not doing that, for three reasons. The message ranges on the Codex pricing page are themselves ranges, 15 to 160 for one model on Plus, because consumption already varies by task. A typical message length for an agentic session is not a number anyone outside OpenAI can defend. And OpenAI's only quantitative statement about December is an expectation about reset frequency, not a figure.

One thing OpenAI has said that bears on the gap: in the Pro tiers article, it notes that a larger allowance does not guarantee twice as many completed tasks, because the amount of work completed still depends on the model, task and settings. That sentence is the right frame for December too. The allowance will be a budget of tokens, and what it buys will depend on how the tokens are spent.

For the record, this is what the Codex pricing page shows on November 5, 2026 for Plus and Standard Business: local-message ranges per model per five-hour period (Pro plans currently have no five-hour limit), with a note that local messages and cloud chats already share the plan's usage allowance and that weekly limits may also apply. It is the baseline the December change replaces, and the ranges are OpenAI's estimates, not promises.

Read the spread in each row and the December change looks less like a new idea than a renaming of one already in force. A range of 5 to 45 messages for the same plan and model is a token-shaped limit described in message units. What December does is make the unit honest and merge the two products' pools. How the pools were split and re-split during the GPT-5.6 rollout is in our GPT-5.6 week-one usage pools report.

Each step below rests on a sentence OpenAI has published, not on an assumed allowance. They are ordered from the one that costs nothing to the one that costs money.

For a team rather than an individual, the budgeting question is the one this change forces into the open: whether seats should be sized by headcount or by the agentic work each person runs. Our report on ChatGPT Work covers what that agent does; our post on Sign in with ChatGPT covers the other direction, where third-party apps spend a user's plan allowance. Both draw from the same pool after December, which is a reason to write the policy down now. Sizing that policy against real usage is part of what our AI transformation engagements do.