Most of your team no longer needs a ChatGPT seat

You can put ChatGPT in front of everyone on your payroll on Monday morning, for nothing, and the handful of people who genuinely lean on it will no longer hit a wall at two in the afternoon and quietly go back to doing the job the long way. Until this week that was a twenty dollar per seat problem. For a forty-person shop where thirty people would use it lightly and ten would use it hard, the honest version of that budget landed somewhere north of eight thousand dollars a year, and most of that money was buying away a rate limit rather than buying a capability. OpenAI just removed the rate limit.

What actually changed

OpenAI announced on Thursday that free accounts move to GPT-5.6 Luna as their default model this week, and that starting next week those same accounts get unlimited text chats along with a Think button that hands harder questions more reasoning time. Unlimited is qualified, and the qualification matters: it covers text, subject to abuse guardrails, and the limits stay in place for file uploads, image generation, and the other tools. The pricing page is already live with the new language, and it is worth reading the free column closely, because everything on it that is not plain text conversation still says "limited."

Paying customers got something too. The version of GPT-5.6 Sol that runs the chat box for Plus and Pro was retuned to give shorter, more direct answers and to stop padding, and there is now a slider for how much thinking you want on a given question. OpenAI also published an accuracy claim that is more interesting than the usual benchmark: on an internal set of financial, medical, and legal prompts that required specific factual detail, responses containing at least one factual error were about 62 percent less common with Luna and 68 percent less common with Sol than with the older GPT-5.5 Instant. That is a claim about the model people get by default, not about a top-end configuration nobody uses.

One line in that announcement is easy to skim past and should not be. The updated Sol is only in the Chat experience. The version powering ChatGPT Work and Codex, which is to say the version doing the multi-step agent jobs, was not touched by this release.

The rate limit was the whole problem

Here is the thing that anyone who has tried to roll a tool out to a team already knows in their bones. Adoption does not die because people are lazy or scared or need another lunch and learn. It dies at the exact moment the tool becomes unavailable.

Picture the sequence, because it is always the same sequence. Somebody on your team finally tries ChatGPT for a real task instead of a party trick. Rewriting a warranty explanation into something a customer can read. Turning four paragraphs of a supplier email into a decision. It works. It works well enough that they do it again an hour later, and now a habit is three or four repetitions into forming, which is the fragile part. Then the free account tells them they have hit their limit and to come back later, or drops them onto a weaker model without making it obvious. The answer they get next is worse. They conclude the tool is inconsistent, which is a rational conclusion from the evidence in front of them, and the habit dies right there. They never tell you. You find out two months later when you open the seat dashboard.

The industry answer to that has been: buy them a seat. Twenty dollars a month per person, times everyone who might touch it, because you cannot predict in advance which three of your forty people are going to turn into the ones who use it for everything. So the rep sizes it at forty seats, offers you a volume conversation, and you write a number into the budget that is mostly insurance against a usage cap.

That advice is now wrong for most of the people on that list. Not wrong for everyone, and the difference is the useful part.

Who still needs a paid seat, honestly

Run your headcount through three questions and the answer falls out.

The first group is everyone whose AI use is a conversation. Drafting, rewriting, explaining, summarizing text that gets pasted in, thinking out loud about a decision, translating a message for a customer, working out how to phrase the hard email. That is the majority of the actual usage in almost every small company, and as of next week it costs nothing and does not run out. These people do not need a seat. They needed one last month.

The second group is anyone whose work goes through a file. Uploading the fifty-page contract, dropping in a spreadsheet of last quarter's job data and asking what changed, feeding it the PDF the insurer sent. The free tier caps uploads, caps memory and context, and caps deep research. If somebody's real workflow starts with dragging a document into the window, the free tier will frustrate them into abandoning it, and twenty dollars is a rounding error against the hours that person spends inside documents.

The third group is anyone building something. If a person on your team is using ChatGPT Work to run a multi-step job, or Codex to ship an internal tool, they are on the paid track and this announcement barely touches them. Free access to Work is listed as limited on desktop only, and the model change did not reach Work anyway.

For a forty-person company that had been quoted forty seats, the honest count after this change is usually somewhere between five and twelve. That is the story. Not a model release, not a benchmark, a line item that got smaller because the thing you were actually buying became free.

Why the accuracy number matters more than it looks

There is a second-order effect here that is easy to miss and is arguably bigger than the money.

The reason a lot of owners have been careful about pushing AI down to the whole team is not cost. It is that the person answering a customer's warranty question at four in the afternoon is not going to fact-check the answer, and a confidently wrong response about a policy or a price or a deadline goes out under your company's name. That fear was well calibrated. It was the correct instinct.

Cutting the factual-error rate on exactly the categories where it hurts, money, health, and rules, changes the risk profile of a broad rollout more than it changes anything for a power user. The power user was already catching errors, because they know what a wrong answer looks like in their own domain. The occasional user was not. And the occasional user is precisely who now has unmetered access.

That cuts in both directions and you should hold both. Fewer errors on a much wider base of usage can still produce more total errors reaching customers, because the base grew faster than the rate dropped. Sixty-two percent fewer mistakes is not zero mistakes, and it is measured on OpenAI's own evaluation, on prompts OpenAI chose. It is directionally believable and it is not a warranty.

The honest take

Four things the announcement does not lead with.

Free is ad-supported, and has been since February. OpenAI's own advertising post is clear that ads run on the free and Go tiers in the US and that Plus, Pro, Business, and Enterprise stay ad-free. The company commits that ads do not influence answers, that conversations stay private from advertisers, and that data is not sold, and there is no particular reason to doubt any of that today. But understand what you are choosing when you deploy free accounts across a team: you are putting an advertising-funded product in the workflow, sponsored placements will sit under the answers, and the thing that makes them relevant is what your people are currently talking about. For most shops that is a shrug. If you are in a business where the topic of a conversation is itself sensitive, sit with it a minute longer than a shrug.

The bigger operational gap is that free accounts are personal accounts, which means there is no admin console standing behind them. No shared workspace, no central data controls, no visibility into who has what turned on, nothing to switch off when somebody leaves. Every person manages their own settings, including the setting that governs whether their conversations can be used to improve the models. On a Business or Enterprise workspace the owner sets that once for everybody. On forty free accounts, the answer is forty separate personal decisions you cannot see. That is the real reason a regulated shop, or anyone handling client data under contract, still pays. It has never been about message caps.

Nothing about this is live yet in the way that matters. The model default is described as rolling out this week and the unlimited chats and Think button as starting next week, which means the correct move today is not to cancel a renewal. Verify it on one account, watch it for a few days, and check that the people who are supposed to be on Luna actually are.

And the meter did not disappear from the world, it moved. Anyone running ChatGPT Business already found that out this month, when the seat fee stopped covering the good features and started drawing them from a credit pool the owner tops up. The free tier going unlimited on text and the paid tiers getting metered on everything else are the same strategy viewed from opposite ends. The cheap thing is now free and the expensive thing is now measured. Plan your budget around what your team actually does rather than around a seat count, because the seat is no longer where the money is.

What this actually costs you

Nothing, on the invoice. Something, on the org chart.

Rolling out free accounts across a team is a decision to accept less control in exchange for much more usage, and both halves of that are real. You will get people using the tool who never would have gotten a seat approved, which is where the value in this whole category has always been hiding. You will also give up the one clean lever a workspace gives you, which is the ability to set a policy once instead of asking forty people to set it themselves.

The reasonable version is neither purist option. Put free accounts in front of everyone and stop treating access as the scarce resource. Keep paid seats for the people whose work runs through files, and for the small number building things. Then write down, in one paragraph, what your team is not allowed to paste into a chat window, and say it out loud to the whole company rather than burying it in a policy document nobody opens. That paragraph is now doing the job the seat licence used to pretend it was doing.

For two years the honest advice to a small operator was that AI access costs about twenty dollars a head and you should be selective about the heads. That advice expired on Thursday, and the interesting question is no longer who deserves a seat but who on your team you have quietly never bothered to tell.