ChatGPT Enterprise pricing: the full breakdown

“OpenAI wouldn’t give us a price without a sales call. What is ChatGPT Enterprise actually going to cost?”

The short answer: there is no list price — ChatGPT Enterprise is negotiated per contract. As of November 2026, public benchmarks put most deals at $50–60 per user per month on a 150‑seat minimum and an annual commitment, so the entry point is roughly $100K a year. The smallest tiers price at the top of that band; larger fleets negotiate toward $40. The numbers below carry dates and sources wherever they’re public.

What does OpenAI actually publish?

Less than you’d expect. From OpenAI’s pricing page (accessed November 2026):

Two details on that page and in OpenAI’s help center matter more than the table. Both workspace plans can buy credits for usage beyond the included limits. OpenAI calls this flexible pricing, and it means the per‑seat fee is a floor once heavy users hit rate limits. Enterprise agreements can now also bill usage by tokens on top of seat fees. The subscription also does not include API usage, which surprises engineering teams often enough to be worth saying twice. Business workspaces can claim a promotional allowance of up to $500 a month in API credits; Enterprise is not eligible for it.

What do companies actually pay?

OpenAI keeps Enterprise pricing off the page, but enough contracts have passed through procurement advisories to establish a band. Benchmarks published by one licensing advisory report, for 2026 negotiations on annual or multi‑year terms:

SpendHound’s buyer data puts the average annual OpenAI contract near $416K for companies with more than 1,000 employees, and near $55K for companies with 50 to 1,000 (updated October 2026).

These are indicative benchmarks, not a rate card. What moves your number: seat count, term length, whether you commit to seat growth, the compliance package you need, and, bluntly, whatever competing quote is sitting on the table when you negotiate.

The per‑seat math to run before signing

Start with the multiplication nobody enjoys: 1,000 seats at the reported $48–55 band is $576K–660K a year. One mid‑size enterprise we spoke with was quoted roughly $500K a year as the minimum to roll a frontier assistant out company‑wide.

Then put the skew next to it. Usage on any enterprise AI platform is lopsided: a small group of power users generates most of the traffic, and a long tail signs in a few times a month. Per‑seat pricing bills both ends identically. Even OpenAI’s newer token‑based Enterprise agreements keep any contracted seat fee as the base, so the light‑usage seats still pay it. The power users are a bargain at $60. The tail is not — those seats cost $660–720 a year each whether they’re used daily or twice a quarter.

A flat per‑seat fee is a bet that your whole company will use AI uniformly. No company does.

Per‑seat does buy something real: a predictable line item. Procurement can budget it, finance can amortize it, nobody gets a surprise invoice. The honest framing is that you’re paying a premium for that flatness, and the premium is the unused capacity of every light‑usage seat.

What you’re paying for (and it’s real)

The Enterprise tier is not padding. Per OpenAI’s published feature list: SCIM provisioning, Enterprise Key Management, role‑based access controls, a compliance API, IP allowlisting, data residency across ten regions, SOC 2 Type 2 plus the ISO 27001 family, and 24/7 priority support with SLAs. OpenAI states it doesn’t train on business data by default.

If your requirement is a managed assistant with mature enterprise controls, and your security posture allows prompts and chat history to live in a vendor’s cloud, it’s a strong product. That second clause is the real decision point, and it has nothing to do with price: for a regulated organization, where the data lives ends the conversation before the per‑seat rate starts it.

What to ask in the sales call

Since the price is negotiated, the call is where the real terms surface. Five questions worth arriving with:

1. What’s the renewal uplift? Year‑one pricing often reflects a land‑and‑expand discount. Ask for a written cap on the year‑two increase; benchmarks are useless if the renewal resets them.
2. What does the seat commitment actually commit us to? A lower per‑seat rate priced against committed growth means you owe those seats whether adoption shows up or not. Model the contract at your realistic adoption curve, not the optimistic one.
3. Credits or tokens? OpenAI now sells Enterprise on credit‑based or token‑based agreements. Ask which one the quote uses, and get the credit or token rates in the quote, not discovered in month three.
4. Is the API in scope? API usage is invoiced separately, though newer agreements can count it toward one shared commitment. Ask how developer access is expected to work and whose budget it lands on.
5. Which of the compliance features are in this quote? Data residency region, Enterprise Key Management, the compliance API: confirm what the quoted tier includes versus what’s an add‑on conversation.

None of this is adversarial. It’s the same diligence you’d apply to any six‑figure annual commitment; the difference with AI seats is that usage data to sanity‑check the commitment usually doesn’t exist yet inside the buying organization.

When per‑seat stops making sense

The alternative model is consumption: models are called through your own provider accounts and billed by the tokens used. OpenAI has moved partway there, with token‑based billing on eligible Enterprise agreements, but usage still runs through OpenAI’s cloud and sits on top of any seat fee. With no seat fee underneath, a seat that sends ten prompts a month costs what ten prompts cost. The trade is that consumption needs governance to be safe — someone has to answer “who spent this” and stop a runaway script before the invoice does.

That governance layer is exactly what SecondStack ships: no per‑seat licensing, LLM usage billed by upstream providers to accounts you control, and per‑user, per‑team, and per‑key budgets with hard cutoffs so the spend curve can’t get away from you. The same virtual keys govern chat, IDEs, and agents through one LLM gateway — and the whole stack is self‑hosted, so the data‑residency question disappears rather than getting negotiated.

In our view, per‑seat pricing fits tools everyone uses in roughly the same amount. Enterprise AI isn’t one of them. Price the tail before you sign for it.

If you’re holding a per‑seat quote right now, run both numbers: the quote, and your actual expected usage priced as consumption. We’ll help with the second one — write to [email protected].