Anthropic ends the discount the moment a customer burns through its committed tokens. OpenAI, according to people familiar with its deals, keeps the old price running for two more months. Named accounts are already switching sides.
What The Information Reported About Anthropic's Discount Policy
The Information reported on October 24 that Anthropic tightened how discounts work for large enterprise customers. Once a client uses up the tokens in its contract commitment, the discount stops. Not at renewal. Right there, mid-contract.
The reporting comes from managers at three software companies that buy from both Anthropic and OpenAI. Neither vendor has published these terms, so everything below carries the same caveat: this is reported behavior, not a public price sheet.
The OpenAI side is different. According to a person close to the company, OpenAI lets discounted customers keep their rate for the rest of the month plus one more month, then moves them to list price.
One vendor treats a discount as a temporary bridge. The other treats it as a runway. That difference is doing a lot of work in enterprise sales conversations right now.
How a Token Commit Works and Where the Price Cliff Sits
Enterprise AI contracts rarely publish a per-token price and leave it there. They sell a commitment: you promise to spend a set amount over twelve months, and in exchange you get a rate well below list.
The structure exists because buyers want budget certainty and vendors want revenue they can forecast. Fine in principle.
The problem is what happens at the edge. Commitments get sized on last year's usage, and in a market where one product launch can double consumption inside a single quarter, that forecast is stale before the ink dries. A team that planned for 400 million tokens and hit 600 million is now paying list price on the overflow. Not a gradual increase. A cliff.
That's the detail that makes procurement teams nervous, and it's exactly where a competing bid gets easy to write.
OpenAI's Counteroffer: An Extra Month at the Old Price
The reported OpenAI concession sounds small: one extra month of buffer after the commitment runs out. In a negotiation, it's worth more than the percentage suggests.
Here's why. A price cliff forces an emergency decision, and emergency decisions favor the incumbent who can ship a new contract in a week. A buffer turns the emergency into a scheduled negotiation. The buyer gets time to benchmark both models on real traffic, which is the thing most enterprise teams never do properly.
It also shifts who absorbs the risk of a bad forecast. Under Anthropic's reported policy, the customer absorbs the whole overage. Under OpenAI's, the vendor eats a month.
Rippling Shifted 36 Points of Coding Work to OpenAI
The clearest evidence in the reporting comes from Rippling. CEO Parker Conrad told The Information that OpenAI's share of the company's coding work went from 16% to 52%. Over the same period, Anthropic's share fell from 57% to 18%.
That's not a small pilot. Rippling builds HR and payroll software, and coding assistants sit at the center of its engineering output.
Read those numbers as a customer interview, not an audit. One company's internal split doesn't prove a market trend, and Conrad has a reason to describe a decisive move.
Still, the direction is hard to ignore. A 36-point swing in one account is the kind of story that gets repeated in every procurement meeting for a quarter. That's a buying signal, not a mystery.
Replit and DataCamp: The Deals Behind the Discounts
Replit signed a contract reported at more than $100 million a year, and switched after OpenAI offered a discount above 30%. That's a big number attached to a single deal, and the reporting doesn't break down which models the workload landed on.
DataCamp CEO Jonathan Cornelissen put the saving differently. He said OpenAI's pricing cut his costs by 80% to 90%. Treat that as one customer's account of his own bill, not a market rate. It won't apply to a 40-person startup.
The most aggressive example in the piece involves a US bank with more than $7 billion in assets: free ChatGPT Enterprise seats doubled to 500 people, at under $7 per seat against a standard price above $20.
Three data points, three different shapes of discount. Vendors negotiate by account size, and the published number never survives contact with a real buyer.
No rate card anywhere.
What OpenAI's Enterprise Pricing Costs in Gross Margin
Discounting to win accounts is an old move, and a price war in enterprise AI is a strange one to watch, because the list price on both sides is a fiction that only small customers pay.
OpenAI has two ways to make a discount stick: charge less than the model costs to run, or serve it more cheaply. It has spent the year on the second, shipping more efficient model lines in Sol and Luna, which lowers the cost of a token even at a lower price. Real price cuts are sustainable. Discounted list prices aren't, and a competitor willing to hand out two free months can undercut them any week.
The competitive pressure runs both ways. DeepSeek and Kimi sell close-enough capability at a fraction of US list prices, and every enterprise buyer now knows it. Neither OpenAI nor Anthropic can hold a premium purely on brand.
Anthropic's IPO Math Turns Discount Discipline Into a Story
Anthropic is reportedly preparing a November IPO, with third-quarter revenue around $16.8 billion, and growth that slowed between the second and third quarters according to an investor memo cited in the coverage.
That context explains the discount discipline. A company heading into a public listing wants clean margins and predictable revenue, and every point of discount given away is a point an investor will ask about.
The tension is obvious. Tightening discounts protects the margin story and hands accounts to the competitor with the looser policy. Loosening them wins the account and weakens the number investors are pricing.
Which way that trade goes is the thing to watch. If Anthropic holds the line through the listing, the reported churn is a cost of going public. If it quietly reopens negotiations in December, the discipline was a quarter-long posture.
What to Ask Before You Sign a Token Commit
If your team is negotiating one of these contracts, the discount percentage is the least interesting line in it. Ask four questions instead.
What happens to the rate after the commitment is exhausted, and does the vendor notify you before the cliff or after? Can unused tokens roll over, and for how long? Is the rate locked for the term, or does it track list price? And what does switching cost in engineering time, since the real lock-in is a codebase wired to one API.
Get those answers in writing. Then size the commitment at 60% to 70% of forecast usage, because the downside of undershooting is a negotiation and the downside of overshooting is paying for tokens you never send.
One more practical note. Both vendors are running the same playbook in opposite directions right now, which means your bargaining power is temporary. Use it at the next renewal, not the one after.
