From December 2026, ChatGPT and Codex will draw on one shared usage allowance for Plus, Pro and Business subscribers. OpenAI explained the change in a help centre article updated on 31 October, and The New Stack reports that OpenAI declined to say how big the combined allowance will be. If your team uses both, the size matters less than you think. What matters is who gets to spend it.
What is changing?
Today, ChatGPT and Codex have their own limits. From December, eligible use of both comes out of one pool.
- Who it affects. Plus, Pro and Business plans.
- What does not count. Ordinary chats with the Instant model stay outside the shared allowance.
- Where to see it. Remaining allowance and reset times show in Settings.
- What happens at the limit. You wait for the reset, move to a higher plan, or buy credits for the features that support them.
What OpenAI has not published is the size of the pool. So nobody can tell you yet whether your team will notice.
Why does sharing the pool matter?
Because chat and coding agents use tokens very differently.
A question in ChatGPT is a few thousand tokens. A Codex task that reads a repository, runs tests and tries three fixes can be hundreds of thousands. When both come from one allowance, a long afternoon of agent work can leave nothing for the quick questions. The person who uses Codex hardest sets the limit for their own ChatGPT, and on some plans that is a surprise you discover at the worst moment, in the middle of a task.
That is not a reason to be angry with OpenAI. A single pool is simpler to explain and probably fairer on cost. It is a reason to stop treating these subscriptions as "unlimited for practical purposes".
What should a team do before December?
Know who your heavy users are
On a Business plan, look at who uses Codex daily and for what. Long autonomous runs on large repositories are the ones that will drain the pool. Talk to those people first. They will hit the limit before anyone else.
Move heavy, repeatable work to the API
If someone runs the same agent job every day, such as a nightly refactor, a test-fixing loop or a dependency upgrade, that work belongs on the API with its own key and a spending cap. You get a clear bill per job and a hard stop, instead of a shared pool that runs out without warning. We wrote about pricing those workloads in GPT-6.1 Terra halves the API price.
Decide on credits now, not at 4pm on a Friday
Buying credits at the limit will be easy, which is the point. Decide in advance who may buy them and up to what amount each month. Otherwise the decision gets made one tired click at a time.
Keep the cheap path cheap
Instant chats do not count. For quick questions, that is the right model anyway. Tell people. A lot of the pool gets spent on questions that never needed a reasoning model.
What would I watch in December?
The first two weeks. Check the usage page every few days for the people you flagged. If they hit the limit, look at what they were doing as well as how much. Usually one or two workflows account for most of the spend, and those are the ones to move or redesign.
A limit you plan for is a budget. A limit you find by surprise stops work. Use December to plan for it.
Where Alongside fits
Deciding which AI work runs on subscriptions and which runs on the API, with cost controls for each, is the kind of work covered by Alongside's technical advisory.
