The “up to 90 percent” in this weekend’s headlines is one customer’s cost reduction, not a discount anyone was offered. Read the reporting closely and the more useful number is 15: the discount Anthropic reportedly gives on large commits, and one that vanishes the day a customer burns through the tokens it committed to buy. That contract mechanic, not the rate on the first page, is what cost Anthropic one of its five biggest customers.
The story comes from The Information’s October 25 report on OpenAI winning coding workloads away from Anthropic, summarized by The Decoder and Runtime Wire. It builds on a September 28 report from the same outlet about how each lab handles the end of a commit. Taken together, they are the clearest look yet at what enterprise AI contracts actually contain, and they say more about what to negotiate than about who is winning.
What the reported deals actually say
Four customers are named, and each one tells you something different.
Rippling. CEO Parker Conrad’s figures, as reported by The Decoder, show OpenAI’s share of the company’s coding workload rising from 16% to 52% while Anthropic’s fell from 57% to 18%. That is a split, not a switch. Rippling still runs both.
Replit. OpenAI reportedly offered more than 30% off a contract worth over $100 million a year, and Replit moved most of its AI spend. Per Runtime Wire, Replit, OpenAI and Anthropic have not publicly confirmed the financial terms. Onstage at Disrupt on October 16, CEO Amjad Masad said “it was a decision based on both cost, partnership, and performance,” and he later told TechCrunch that “Replit remains model agnostic.” People familiar with the partnership disputed his account to TechCrunch, and Replit’s own blog cited Claude Fable 5 and 5.1 in production as recently as late September.
DataCamp. This is where the 90 comes from. Its CEO said OpenAI’s pricing cut DataCamp’s costs by 80 to 90%, according to The Decoder. That figure blends a negotiated discount with a move to cheaper models. It is not a rate card.
An unnamed US bank with over $7 billion in assets saw its ChatGPT seats doubled to 500 at no extra charge, paying under $7 per user against the usual $20 or more, per the same report.
None of the four is a verified contract. All of it is secondhand, from executives with their own reasons to talk. Treat the figures as signals of how hard OpenAI is pushing, not as a benchmark to quote at your account rep.
The cliff at the end of the commit
The detail that matters most for buyers is in the September report. Managers at three software companies that buy from both labs told The Information that Anthropic’s discount ends once a customer uses up the tokens it committed to, as The Next Web and Dealroom summarized. Here is how the three approaches compare, based on that reporting:
| Vendor | Discount on large commits | When the committed tokens run out |
|---|---|---|
| Anthropic | About 15% off list (reported) | Discount ends; renegotiate or pay list |
| OpenAI | Not disclosed | Rest of the month plus one more month to sign a new deal, then list |
| AWS, Microsoft, Google | Varies | Discount usually continues on overage for the rest of the term |
Anthropic declined to comment. The hyperscaler line comes from Jeff Muscarella of NPIFinancial, quoted in The Next Web.
The Replit story is that cliff in action. According to the October report as summarized by Michael Parekh, Anthropic would not extend a discount it had previously given Replit unless Replit committed to spend more. OpenAI’s offer arrived instead. Moving over took hours of code changes and a few days for Replit’s products, which the report says was faster than negotiating the earlier Anthropic contract.
That last detail is the leverage. When moving a workload between frontier APIs takes days, the commit cliff stops being a reason to sign a bigger commit and becomes a reason to leave.
I have been on the buying side of this kind of term for most of my 20 years in IT operations at a large telecom, mostly on carrier circuits and enterprise software licenses. The vendors whose contracts I trusted were never the ones with the best first-year rate. They were the ones whose overage clause did not punish us for growing faster than forecast. A minimum commit is a bet on your own usage; a hard cap at the commit means the vendor collects if you guess low and reprices you if you guess high. For AI spend, where agent workloads routinely blow past their forecasts, guessing right is the exception.
Share of wallet is the clause to read twice
The quieter finding in the September report is that both labs have started adding “share of wallet” language to large discount deals: terms that require a customer to send most of its AI budget to one platform. CodeRabbit CEO Harjot Gill and another manager described the clauses to The Information, per The Next Web. Consultants quoted by Dealroom said neither lab can verify compliance.
Unverifiable today is not the same as harmless. A share-of-wallet clause is an exclusivity term with a discount attached, and it cuts against the strategy most enterprise teams have settled on: route each task to the cheapest model that passes your evals and keep a second provider qualified. If you route traffic through a model router or tier work across vendors the way a cost map suggests, a clause promising 70% of spend to one lab quietly overrides your architecture. Ask how “AI budget” is defined, whether spend through Bedrock, Vertex or a router counts, and what happens to the discount if you fall short.
What a discount is worth when list prices fall by half
The other reason the headline percentages mislead is that the list prices underneath them are moving faster than any contract. Max Woolf’s tally of 2026 price cuts shows GPT-6 Sol launching on September 22 at $2 input and $10 output per million tokens, half the discounted GPT-5.6 Sol price, and Claude Opus 5.5 launching the same day at $4/$20, down from Opus 5’s $5/$25.
Run the arithmetic on that (our own math, not a reported deal). A 15% discount on Opus 5 works out to $4.25 input and $21.25 output. Opus 5.5 at plain list price is cheaper on both. If your discount is written against a specific model’s rate card, it may already be worth less than nothing. If it is written as a percentage off whatever the current list price is, it moves with the market. Which of those you signed is a sentence in the order form, and most teams have never checked.
The money behind the offers matters too. Anthropic spent $5.6 billion on compute in Q2 against $11.5 billion in revenue and has told partners it earns roughly a 75% gross margin on existing models, while OpenAI has projected compute costs well above its revenue this year, per Parekh’s summary. The Decoder describes OpenAI as partly buying market share with its own margin. Subsidized pricing ahead of an IPO is real money while it lasts, and it is also the price most likely to change at renewal.
So what does a buyer take from a week of leaked deal terms? Not “switch to whoever discounts most.” The terms worth fighting for are the ones these reports show actually deciding outcomes:
- Overage pricing. Keep the discount on usage above the commit for the full term, the way the hyperscalers usually do. If the vendor will not, get a grace period in writing.
- Price-cut pass-through. Write the discount as a percentage off current list, so a September repricing like the Opus 5.5 and GPT-6 Sol cuts lowers your bill instead of stranding it.
- No share-of-wallet term, or a narrow one that excludes routed and cloud-marketplace spend and carries no clawback.
- A renewal ceiling, so a first-year subsidy does not become a second-year surprise.
- Portability you have tested. Replit’s switch took days because its stack was already model agnostic. Rerun your own eval suite against a second provider every quarter, so the threat to leave is one you can carry out.
The labs are competing hard for exactly the kind of high-volume coding and agent workload most enterprise AI budgets now go to. That competition is the best leverage buyers have had since these APIs launched, and it shows up in the overage clause, not the headline rate.