TalkWhat a seat means when the worker is software
For thirty years a seat was the most honest unit in software. One person, one login, one price. The word carried an assumption so obvious nobody wrote it down: a person can only sit in one chair, type so fast, and want so much. AI assistants have kept the word and removed the assumption. The chair is still there; what sits in it is increasingly a process the person started. Both major vendors' pricing pages, read on October 10, 2026, show them working out in public what to charge for that, and this essay is about what a small company should put on the purchase order instead of a headcount.
The seat used to describe you; now it describes the vendor
A seat price was a bet the vendor made about an average person. Some users would barely log in, some would live in the product, and the price was set where the heavy ones were paid for by the light ones. The customer never saw the bet. They saw a number per head, multiplied it by the org chart, and were done.
The bet worked because the worst case was bounded by a body. Nobody could cost the vendor a hundred times the average, so the vendor never had to say what the average was.
An assistant that can summarise forty documents in the time it takes to ask, or run a coding task while its owner is at lunch, breaks the bound. The expensive user is no longer a fast typist but someone who has learned to delegate, and delegation has no ceiling. The seat price stops describing the customer's average person and starts describing the vendor's average cost.
Both vendors split the chair in two, at the same prices
The evidence is on the two pricing pages. OpenAI's business pricing page, checked October 10, 2026, lists a Standard seat at $20 a month billed annually or $25 monthly and a Premium seat at $100 or $125, described as five times the usage, with no 5-hour limit. Anthropic's page lists Claude Team Standard at $20 or $25 and Premium at $100 or $125, five times Standard.
The numbers are identical, which is interesting, and the structure is identical, which is more so. Both vendors concluded that one seat price could no longer cover the range of appetites inside a single small company, and both drew the line at the same place: a $20 chair for a person who chats, and a $100 chair for a person who delegates.
Anthropic's Team help article puts numbers on the halves: a Standard seat has 1.25 times the Pro plan's per-session allowance and a Premium seat has 6.25 times. OpenAI's only hard statement is that the Premium seat has no 5-hour limit, which tells you by subtraction what the Standard seat has. The ChatGPT and Claude subject pages quote both in full.
Two seats, same person, same chair, five-fold price difference. The thing being priced was never the chair.
Then both vendors invented a seat that costs nothing until it works
The split is the first step. The second is already on the pages. OpenAI's Enterprise help article, dated 2026, describes a Codex seat with no fixed cost per user per month, billed entirely on usage through workspace credits. A person holds the seat; it costs nothing; the moment the software in it does work, the meter runs.
Anthropic has gone further on its Enterprise plan. The pricing page lists $20 per seat per month billed annually, and the Enterprise help article says the seat fee covers access only, with all usage billed separately at API rates, no per-seat limits and no included allowance. The seat is a key to the building. Everything done inside is charged by the token.
Below Enterprise, both vendors have added a valve pointing the same way: usage credits at API rates on Claude's paid plans, purchasable credits on ChatGPT Business and Enterprise. The seat sets the included portion; the credits are the metered portion; the line between them is what the vendor adjusts.
Read in sequence, the direction is not subtle. One seat became two seats. Two seats became a seat plus a meter. On the most expensive plans, the seat is now a formality and the meter is the product.
Why the vendors cannot go back
It is tempting to read this as greed, and that reading misses the mechanism. The vendors did not choose to make the seat ambiguous; the product did. A tool that works while you are elsewhere consumes resources in proportion to how much you trust it, not in proportion to how many of you there are.
Anthropic's pricing page says this almost in those words: usage depends on the length and complexity of conversations, the model, the features and the effort level selected. Both vendors are explaining that the cost of a seat depends on what the occupant asks it to carry.
Once that is true, a flat seat price is a subsidy flowing from light users to heavy ones, and heavy users are the ones most likely to discover how heavy they can be. The split seat and the meter are not a strategy; they are the only stable place to stand. We made the same argument about the word at the other end of the scale in Unlimited is a pricing word.
What this does to a small company's budget
The headcount multiplication no longer produces a budget. It produces a floor. Ten Standard seats at $20 a month billed annually is $2,400 a year at either vendor, an editorial calculation from the pricing pages. That is the least the company can spend if everyone stays light.
The actual figure depends on a question the org chart cannot answer: how many of the ten will learn to delegate? Each one who does is a $100 seat, and the pricing pages encourage you to move them rather than let them hit a wall. Two of ten on Premium takes the year to $4,320. Four of twenty takes it to $8,640.
So the budget line should be written differently. Not seats times price, but:
- A floor: the Standard seat count times the annual price, which is the same at both vendors.
- A delegation allowance: a guess at how many people will become heavy within the year, times the gap between $20 and $100. Guess high; delegation spreads.
- A meter reserve: a small sum for usage credits in the months when a project makes somebody heavier than their seat, which both vendors now sell.
- A review date: the quarter in which you look at who is actually on which seat, because the vendors' usage pages will show you and the renewal will not wait.
The primer on AI assistant pricing for small teams works through the arithmetic for two, ten and twenty people, and the comparison shows how little the two vendors' floors differ. Who gets the heavy seats is a question of trust rather than seniority, since the person on the $100 seat is the one authorising software to act on company documents while nobody watches; Who reads what you type is about that.
The strongest objection, and why it does not hold
The objection is that this is temporary: compute gets cheaper every year, so the meter will fold back into a flat seat and a small company should wait it out. Two things on the pages argue against it.
First, OpenAI's Pro tiers article records a Pro 200 plan whose included usage was reduced to reflect, in the vendor's words, increasingly efficient models, with the price held at $200. Cheaper compute did not produce a cheaper flat plan; it produced a smaller allowance at the same price.
Second, the newest models on both pricing pages are the most tightly rationed: Anthropic lists its Fable line at 50 per cent of weekly limits on Premium seats, and OpenAI reserves its Ultrafast variants for the $500 tier. Efficiency lowers the cost of yesterday's model while today's model raises the appetite. As long as there is a better model to want, the meter will have something to measure.
What follows in practice
Stop multiplying heads by $20 and calling it a budget. Write the floor, the delegation allowance, the meter reserve and the review date, and expect the second line to grow. Assign heavy seats by what you are willing to let the software do unsupervised. Read the usage page quarterly; that page, not the pricing page, is where your seat price is now written.
Buy the company plan rather than reimbursing personal ones; the subject pages for ChatGPT and Claude explain why the defaults differ. Before signing anything annual, read The claim that you can always switch, because habits of delegation do not export.
And remember the older argument in The tool is not the workflow. The seat that costs the most is the one in which someone has finally built a workflow worth delegating. That is not a cost to be minimised. It is the only reason to be paying at all.
Sources, checked October 10, 2026: openai.com/business/pricing, openai.com/chatgpt/pricing, OpenAI help center (Enterprise, Pro tiers); claude.com/pricing, Claude help center (Team plan, Enterprise plan).