Deep Dive — ChatGPT Enterprise price: $7 seats and 90% discounts

OpenAI is handing out discounts of 30 to 90 percent to pull large customers off Anthropic, and three named companies have already moved. Rippling flipped its coding workload from Claude to ChatGPT. Replit signed a contract reported at more than $100 million a year. DataCamp's chief executive put his savings at 80 to 90 percent. In one US bank, 500 employees got ChatGPT seats at just under $7 a user, where the usual figure is $20 or more.

The reporting comes from The Information, under a headline that is also the analysis: "Anthropic's Margin Is OpenAI's Opportunity," the outlet reported on October 24. That framing is right, and it is the reason this story is bigger than a handful of sales wins. But the most-quoted number in it — 90 percent — is not a rate anyone has been shown receiving. What has been granted, on the reporting, is narrower, conditional, and tied to a second product. And the switch most people will cite as proof, Replit's, began with an expiring discount rather than a benchmark loss.

That distinction matters if you are the one signing the renewal.

What OpenAI is actually handing out

The clearest datapoint is Rippling. Chief executive Parker Conrad said OpenAI's share of the coding work done by his engineers went from 16 percent in May to 52 percent, while Anthropic's fell from 57 percent to 18 percent. Read the two pairs of numbers carefully and one thing is obvious: they do not sum to 100 in either period — 73 percent then, 70 percent now — so the remaining share sits with other models or in-house tooling. This is a shift in the largest slice of one company's coding work, reported by its chief executive. It is not a two-vendor seesaw and it is not a market-wide survey.

Replit is the big one, because of the size of the contract rather than the size of the discount. The reported terms: more than 30 percent off a deal worth over $100 million a year. The reported trigger: Anthropic declined to extend the discount Replit had been receiving unless Replit committed to spending more, and OpenAI's offer arrived after that. Replit chief executive Amjad Masad said on stage at TechCrunch Disrupt on October 16 that the company was no longer working with Anthropic, citing cost, partnership and performance, and adding that he thought ChatGPT was better now. Within days he had qualified it: Replit "remains model agnostic," OpenAI is its preferred partner for powering Replit Agent, and users can still pick Claude. Replit's own model selector documentation, checked this week, still lists seven Claude models across its Power and Max modes, alongside OpenAI's and Kimi K3.

DataCamp is where the 90 percent figure enters, and where it should be handled with care. The grant The Information reported was a 40 percent reduction on models used for internal product development, conditional on DataCamp also using Codex. Chief executive Jonathan Cornelissen estimated his costs could fall 80 to 90 percent if comparable terms were extended to the models inside the products DataCamp sells. He was describing a reduction he was seeking, not one he had signed. The Decoder's October 25 writeup, which is where most of the English coverage starts, flattened that estimate into OpenAI "discounts up to 90 percent."

The bank item is the one to keep, because it is the only place a per-seat price appears: more than $7 billion in assets, its ChatGPT seats doubled to 500 employees at no extra charge, paying just under $7 per user against what the report describes as a usual $20 or more.

Why the list prices tell a different story than the discounts

Here is the part the discount narrative skips. At the tiers agents actually run on, the two labs' published prices are now identical to the cent.

OpenAI's business pricing lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output, and GPT-6 Luna at $0.10 and $0.50, according to OpenAI's price page. Anthropic's list has Claude Sonnet 5.5 at $2 and $10, and Claude Haiku 5.5 at $0.10 and $0.50. Four numbers, two companies, no daylight. The labels differ; the meter does not. Only at the top does the gap open — GPT-6 Astra at $10 input and $50 output against Claude Opus 5.5 at $4 and $20, which makes Anthropic's flagship the cheaper of the two best models, at least on the card.

So the 30-to-90 percent discounts are not being carved out of a rate-card gap, because on the volume tiers there is none to carve. They are being carved out of contract structure: committed-spend thresholds, seat bundles, multi-year terms, and — the mechanism that actually moved Replit — the expiry of a discount the incumbent had been granting. Cache pricing is the other lever. Luna's cached input rate is $0.01 per million tokens, a tenth of its already-cheap input price; for a coding agent that re-sends the same repository context on every turn, that line item can matter more than the headline percentage on a quote.

Some of the discounting is also genuine cost reduction being passed through, and it predates this story. OpenAI cut GPT-5.6 Luna's price by 80 percent on July 30 and took 20 percent off Terra in the same release; a month later it cut Sol's input price by 20 percent and its output price by a third. Newer Sol and Luna versions landed about 50 percent below the models they replaced. A lab that halved its own serving cost twice in three months has room to discount that a lab which did not, does not.

The margin being priced against

Anthropic's own numbers explain why OpenAI is aiming here. In its third-quarter update to potential investors, Anthropic reported $16.8 billion in revenue, roughly 45 percent above the second quarter — down from the more than 140 percent sequential growth it posted in Q2 — plus an adjusted operating margin of 11 percent for the year through Q3, a second straight quarter of adjusted profitability, and $1.70 of revenue for every dollar it spent on compute, The Decoder reported from the investor meeting and the coverage around it.

That $1.70 is the number a competitor reads first. Roughly 70 cents of gross margin on every compute dollar is a negotiating range, and a discount of 30 percent on a large contract is a claim on it. A rival willing to run at or below cost on a logo can spend that margin faster than the incumbent can defend it, because defending it means cutting into the one metric the incumbent's valuation now rests on.

The timing sharpens this. Anthropic's IPO is planned for November, before Thanksgiving, on a private valuation that reached $965 billion in May, with bankers having floated as much as $2 trillion and several investors now expecting expectations to be cut to under $1.5 trillion. A company pricing a listing needs two things at once: revenue growth that still looks steep, and a margin story that survives scrutiny. Discounting defends the first and damages the second. That is the trap in the numbers — the cheapest way to keep a customer this quarter is the most expensive way to explain yourself to an allocator next month.

OpenAI has the opposite problem and can afford the trade. Its annual recurring revenue was reported near $70 billion in late September by Axios, and it has been spending beyond revenue on compute. Buying a default position in coding now is a bet that the default is worth more later than the discount costs today — the same logic that built every platform business, with a compute bill attached.

What the discount is actually buying: the default

The most underrated fact in this story is that Replit's documentation still lists Claude. That is not a contradiction of Masad's statement; it is what a customer default looks like. Replit stopped routing to Anthropic as its primary path and made OpenAI the preferred partner behind Replit Agent, while leaving the model selector open for the users who want something else. No rip-and-replace, no migration announcement, no benchmark graph — a default flipped, and the default is where the volume goes, because developers do not re-pick a model on every turn.

That is also why the trigger being a discount expiry matters more than the size of the discount. An incumbent that prices a large customer at a special rate, then asks for a bigger commitment to keep it, has converted a pricing decision into a procurement event. Procurement events invite competing bids. Anthropic's instrument in the Replit case was a discount too — the switch happened because one lab would not extend and the other would — which makes this a renegotiation with switching costs, not a capability verdict.

There is a counter-reading worth stating plainly: if discounts of this size were reshaping the market, Anthropic's third quarter should have shown it, and Q3 came in before the loudest of these deals closed. The Replit announcement was August; Rippling's share shift is measured from May; Anthropic's quarter ended in September. So the reported revenue largely precedes the discount wave rather than reflecting it. That cuts both ways. It means the story is not yet visible in the numbers, and it means the quarter the company will be asked about during an IPO roadshow is the first one where it might be.

There is a second, unrelated headwind stacked on the same quarter. Anthropic estimates 5 to 7 percent of its revenue comes from distillation — other companies extracting its outputs to train rivals — and it has now deployed countermeasures that cut reasoning-token theft by up to 99 percent. Shutting that off removes revenue. An investor looking at Q4 will have to separate a deliberate anti-distillation decline from a competitive one, and the company has not given anyone the split.

What this means if you are the buyer

Three rules come out of this week, and none of them is "take the biggest percentage."

First, price the contract at renewal, not at signature. The Replit case is the template: the number that mattered was not OpenAI's benchmark score or even its list price, it was what happened when a discount came up for renewal. Ask for price protection across the term, a written definition of the committed-spend threshold that unlocks the rate, and an exit if the discount is not renewed. A discount with a one-year fuse is a customer-acquisition cost the vendor is amortizing, and you are the asset being amortized.

Second, measure cost per task, not cost per token. We went through the arithmetic in Haiku 5.5's fine print: same price, three times the bill: Claude Haiku 5.5 and GPT-6 Luna quote identical per-token prices at their top effort settings, and measured per-task cost came in at $0.21 against $0.07, because the Anthropic model burned 162,000 output tokens where the OpenAI model burned 50,000. A 40 percent discount on the more verbose model is still a loss. Run your own traces through both before you take a percentage at face value.

Third, do not confuse a cheap seat with a cheap workload. The bank paying under $7 a user is buying chat seats, and seats are the smallest line in an AI budget once agents are running. The bill that grows is the API meter, and that is the one with the rate card both labs have already cut twice this year.

The floor neither lab controls

The reason neither company can hold a price is that the price of "good enough" keeps falling from outside the duopoly. DeepSeek's V4.1-Flash-1017 landed at 77 percent on DeepSWE for about 20 cents a task, which we covered in detail. Moonshot's Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index — three points behind Claude Opus 5.5 — and lists at $2 and $10 per million tokens with weights promised by November 4. MiniMax's M3.1 Flash put up 42 on the same index at $0.22 a task.

Set those against a US enterprise contract and the discount strategy reads differently. A 90 percent discount is not generosity; it is a window. The lab that wins a coding customer at 40 percent off this quarter is defending against a model that will be downloadable at a tenth of the list price next quarter, and both American labs know it. That is the real reason the price war has no floor: the ceiling on what either can charge is set by weights neither of them publishes.

What to watch

Anthropic's fourth-quarter revenue, because it is the first period that fully contains the discount wave and the first one IPO investors will question. Whether the company discloses an annualized revenue rate alongside it — it declined to give one in the Q3 meeting, after reporting $65 billion in July. The IPO price against that $965 billion mark: a listing below $1.5 trillion would say the market has decided the growth deceleration matters more than the margin.

And watch the shape of the next deal rather than its percentage. If OpenAI starts publishing committed-spend tiers instead of one-off discounts, the war has moved from buying logos to locking them, which is a different and more durable strategy. If Anthropic answers with capacity credits or priority access instead of price, it has decided to defend margin and let some customers go — the choice every incumbent eventually makes, and the one its IPO pricing will reward or punish.

For now, the honest summary is the unglamorous one. Two labs with near-identical rate cards are competing on the only variables left: contract terms, defaults, and who blinks first on a renewal. The 90 percent headline is a good story. The $7 seat and the expiring discount are the parts you can actually take to a negotiation.

If your AI contract came up for renewal this year, what did the vendor actually give up to keep you — and what did it ask for in return? Write to us via our contact page.