ChatGPT Business just got a meter, and it charges for the good part

Your ops lead ran a deep research report on a prospect this morning, the kind that reads forty sources and comes back with a two-page brief. Last month she ran that report whenever she wanted and it cost nothing past the seat you already paid for. Starting now, that same report draws 50 credits out of a shared pool your workspace has to keep filled, and when the pool runs dry the button stops working until somebody with billing access tops it up. The flat seat you budgeted did not go away. It just stopped being the whole bill.

That is the shift OpenAI finished rolling into ChatGPT Business, and the rate card that puts a number on every advanced feature updated in the last day. The seat is still the seat. What changed is that the features people actually reach for when the tool earns its keep are now on a meter, and the meter is a credit pool the owner buys and refills.

What actually changed

The mechanics are laid out in OpenAI's own flexible pricing explainer, and they are worth reading slowly because the summary everyone will repeat gets them wrong.

Nothing was taken away from the plan. Your seat still includes the everyday chat models with virtually unlimited use, file upload, search, canvas, and the connected apps. GPT-5.5 Instant, the model most messages land on, costs zero credits and stays that way. If the whole team does is type questions into the box, nobody ever sees a meter.

The meter lives on the advanced features. Each ChatGPT Business seat now gets a per-seat included allowance for things like deep research, the thinking models, image generation, voice, and agent mode. Cross that per-seat line and, if the workspace has bought credits, the overflow draws from a shared pool everyone pulls against. Run the pool to zero and the feature is blocked, with an in-product button to beg an admin for more. The rate card prices the individual actions in credits: a deep research task is 50 credits, an agent-mode message is 30, a GPT-5.5 Thinking message is 10, a GPT-5.5 Pro message is 50, an image is 5, and a minute of voice is 5. The heavier Excel and Sheets work and the Workspace Agent runs are priced by tokens instead of a flat per-action number, landing somewhere around 5 to 30 credits a task depending on how much text goes in and comes out.

There is a quieter change underneath. As of June 24, new Business workspaces can no longer add Codex seats at all, and only workspaces that already had one before that date get to keep managing them. So the coding-agent side of the plan is being closed to newcomers at the same moment the rest of it goes metered. If you were planning to grow into that, the door is already shut.

Why it matters to a small operator

Here is the part the vendor language walks around. The features that got the meter are not the throwaway ones. They are the reason a small team pays for the plan instead of the twenty-dollar consumer tier. Deep research is what replaces the half-day somebody used to spend building a prospect brief by hand. Agent mode is the thing that clicks through a workflow instead of a person. Voice is the intake call. Image generation is the ad mockup the owner used to farm out. The plan kept the chat box free and put the meter on exactly the work that was doing the replacing.

That matters because of how "included" trains behavior. When a capability feels free, people use it without thinking, and that is the whole point of getting a team onto a tool. You wanted the office manager running deep research on every inbound lead. You wanted the two-person marketing function generating a dozen image options a day. You spent months getting people to build the habit. The meter arrives right after the habit sticks, and now every reflex you worked to install is a draw against a balance. Adoption and spend became the same line on the graph, pointing the same direction.

For budgeting, the seat price is now a floor, not a ceiling. ChatGPT Business runs about twenty dollars per user a month on the annual plan with a two-seat minimum, and that number tells you the smallest the bill can be, not the largest. The largest depends on what the team does inside the tool, and the team does not see the meter while they work. Somebody generating image after image to get a layout right is spending five credits a shot and has no reason to feel it. The person who feels it is you, at the end of the month, reading a usage report.

There is a real operational upside here if you catch it early. The credit and spend-control settings let a workspace owner set monthly credit limits by seat type and per user, cap the whole thing, and turn on usage alerts before a threshold is crossed. That is a genuinely useful lever. It means you can hand the whole team deep research and still know the outside edge of what a bad month costs. But it is off by default. Out of the box, every seat and user has no limit specified, which is to say the tool ships uncapped and it is on you to go bolt the cap on before anyone finds the accelerator.

The honest take

Start with the number OpenAI does not print. The rate card tells you a deep research task is 50 credits. It does not tell you, anywhere public and fixed, what a credit costs in dollars, because the conversion moves with your plan and contract. So the sticker says 50 and you cannot, from the sticker alone, tell whether that is loose change or a real line item. That is a strange place to leave the one number an owner most needs, and it means the only way to know your true cost is to run real usage and read the bill after. Pricing you can only understand in hindsight is pricing that is hard to plan against, and planning against it is the entire job you are being handed.

Then there is the auto-reload. You can set a minimum balance so the workspace tops itself back up to a target automatically when it runs low, which keeps people from getting blocked mid-task. Convenient. The catch is in the fine print of the same setting: leave the monthly recharge limit blank and you have authorized unlimited automatic top-ups against the card on file. A runaway workflow, a well-meaning employee looping an agent, a script somebody left running, and the meter keeps spinning and the card keeps paying, with no ceiling unless you went and set one. Auto-reload with no monthly cap is a standing invitation for a surprise, and it is a box you have to know to check.

The credits themselves are not generous on terms. For Business they are valid twelve months and they are not refundable, save for the narrow cases the law forces or a confirmed account compromise. So the natural defensive move, buy a big pack up front so nobody gets blocked, quietly means committing money you cannot get back to a consumption pattern you have not measured yet. Buy small and people hit walls. Buy big and you are floating OpenAI an interest-free balance you might not burn.

One more, and it is the sort of thing that reads fine until it is your invoice. The rate card notes that the everyday Instant model may, on a request it judges complex, automatically route to the Thinking model, and when that happens you are charged the Thinking rate, 10 credits, logged under a line called gpt-5.5-auto-thinking. Read that plainly. The system decides a message deserves the pricier model, on its own, and bills you for the upgrade you did not ask for. It routes the other way too, dropping simple prompts to a mini model for zero credits, so it is not a pure money grab. But the direction that costs you is a decision the tool makes without you, and it will show up as a category on the bill you did not create.

Who is this genuinely wrong for? The shop that bought ChatGPT Business specifically to put deep research and agent mode in front of a whole team, all day, without thinking about it. That was the pitch and that is now the expensive path. If your usage is heavy and steady on the metered features, you are the customer the meter was built for, and you owe yourself a session with the usage report and the spend caps before month two, not after. The team that mostly lives in the chat box will barely notice any of this. The gap between those two teams is the gap between a flat bill and a variable one, and which side you land on is decided by habits you already spent months building.

What to do about it

Go into billing this week, before the next usage cycle, and set a monthly credit cap per seat and a real number in the auto-reload limit. Then pull one usage report after two weeks and find out what your team actually does when it thinks the tool is free, because that number, not the seat price, is what you are really paying for now.