OpenAI Is Offering Up to 90% Off to Steal Anthropic's Biggest Customers, the Frontier Price War Just Went Private
The list price is no longer the price
For most of this year, the frontier AI price war was fought in public: a lab cut its published per-token rate, rivals matched within days, and anyone could read the new numbers off a pricing page. Reporting published on Sunday, October 25, from The Information, relayed by The Decoder and several other outlets, shows the war has moved somewhere harder to see and harder to answer: into private enterprise contracts, aimed account by account at Anthropic’s largest customers, with discounts reportedly as deep as 90 percent.
The named accounts are what make this more than a rumour. These are not anonymous “sources familiar with the matter” describing a strategy. They are chief executives, on the record about their own companies’ spending, giving before-and-after numbers.
- Rippling. CEO Parker Conrad said OpenAI’s share of the company’s coding workload rose from 16 percent to 52 percent, while Anthropic’s share fell from 57 percent to 18 percent. That is not a pilot or a toe in the water, it is the majority of a large company’s code generation changing hands.
- Replit. The developer platform switched after OpenAI offered more than 30 percent off a contract worth over $100 million a year. Secondary reporting adds the detail that Anthropic had declined to extend Replit’s earlier discount unless it committed to spending more. CEO Amjad Masad said on stage in San Francisco on October 16 that the company was no longer working with Anthropic; he later clarified that Replit “remains model agnostic” and that “OpenAI is now our preferred partner for powering Replit Agent.” OpenAI is now the default there, though Replit’s documentation still lists seven Claude models.
- DataCamp. CEO Jonathan Cornelissen said OpenAI’s pricing cut his company’s costs by 80 to 90 percent. One secondary report puts DataCamp’s specific Codex discount at 40 percent, a reminder that the headline figures bundle different products and deal structures, and should be read as reported, not audited.
- An unnamed US bank. A bank with more than $7 billion in assets reportedly had its ChatGPT seats doubled to 500 employees at no extra charge, and pays under $7 per user against a usual rate of $20 or more. Doubling the seats for free is not a volume discount. It is a land grab measured in headcount.
Buying share, or passing on real savings? Both, in unknown proportions
The easy reading is that OpenAI is buying market share at the expense of its own margins, and The Decoder’s summary says exactly that, in part. But the same reporting carries the other half: OpenAI has also achieved real cost reductions through its newer Sol and Luna model lines, thanks to more efficient models and infrastructure. A discount funded by genuine efficiency is sustainable. A discount funded by investor capital is a countdown.
Buyers cannot easily tell which half they are receiving, and for now they may not care. If your AI bill falls 80 percent, whether your vendor profits on your account feels like next year’s problem, until renewal, when the market has consolidated around whoever survived. That is the classic shape of these fights: the discount is real, and so is the lock-in it purchases. Rippling’s 16-to-52 swing is the sound of switching costs being built in the other direction.
Why Anthropic is exposed, and why the timing is brutal
Three pressures converge on Anthropic at once.
First, the timing. Anthropic is weeks from a reported stock-market listing, this site covered its IPO prospectus on October 2, and the same reporting cycle that carries the discount story also carries an investor memo saying Anthropic’s growth slowed from the second quarter to the third. A growth slowdown plus named customers explaining, in public, how much cheaper the rival got is close to the worst pre-IPO narrative a company can draw. None of the reports included a statement from Anthropic.
Second, Anthropic’s pitch has never been price. It sells capability, safety and enterprise trust, the reasons a customer pays a premium rather than taking the cheapest adequate model. That pitch works while the capability gap is wide. It gets harder each time a cheaper model is good enough for the workload in question, and coding, the workload in every example above, is precisely where “good enough, much cheaper” bites first.
Third, the pressure is not only from OpenAI. Both American labs face a rising floor from China: DeepSeek and Moonshot’s Kimi line keep delivering comparable performance at a fraction of Western list prices, a pattern this site has tracked all month, most recently with Kimi K3.1 Pro at $2/$10 per million tokens. When a credible alternative costs a tenth as much, a premium vendor’s negotiating room shrinks whether or not the customer ever actually switches to the cheap option. The discount OpenAI is offering Anthropic’s customers is, in part, the Chinese price umbrella arriving in an American sales negotiation.
Read this alongside Saturday’s other number
There is a useful piece of context from the same weekend. On October 24, Apollo Global Management’s chief economist Torsten Slok published a note titled “No Signs of AI Payoff Outside the Tech Sector,” finding revenue per worker and profit margins rising for the Magnificent Seven and flat for the other 493 companies in the S&P 500, with mainstream firms starting to treat token spending as a cost centre rather than a growth engine.
Put the two stories together and the discount war stops looking like generosity. If most customers cannot yet point to a payoff from AI spending, price becomes the argument that decides renewals, the only line in the business case unambiguously improving. Vendors are discounting because of weak customer returns: a customer who cannot prove AI made them money will demand that it at least cost less, and move workload to whoever says yes.
The honest caveats
- This is single-source reporting, relayed widely. The underlying account is The Information’s; the specific figures come through summaries by The Decoder and others. The CEOs quoted are named and speaking about their own spending, which strengthens it, but the contracts themselves are private, and neither lab has published the terms.
- “Up to 90 percent” is one customer’s reported cost reduction, not a rate card. DataCamp’s 80-to-90 percent figure may reflect moving workloads to cheaper model tiers as much as a negotiated discount on the same service. Treat the headline number as the top of a range that starts at “more than 30 percent off a nine-figure contract”, which is dramatic enough without the maximum.
- Share of workload is not share of revenue. Rippling moving coding share does not mean Anthropic lost the same share of Rippling’s total AI spend, and Replit still lists Claude models in its documentation. Multi-model is the norm; “poached” accounts are usually rebalanced accounts. The direction is clear even if the final split is not.
What to do with this, this week
- If you buy AI at any scale, renegotiate now. Named nine-figure customers are getting 30 percent or more off, and seat deals are being doubled for free. Whatever you signed six months ago is above the current market. Ask, with a credible alternative in hand, because that is what made these discounts happen.
- Price the renewal, not the discount. Before moving a workload for a steep cut, ask what the rate is in year two, what happens to your pricing if you consolidate onto one vendor, and how portable your prompts, evaluations and tooling are if you need to move back.
- Keep a second model warm. Every company in this story that got a dramatic deal had a real alternative already in production. Multi-model architecture is no longer an engineering preference; it is negotiating leverage with a measurable dollar value.
- If you sell on capability, bring receipts. Anthropic’s problem this weekend is not that its models got worse; it is that its customers talked about price instead of outcomes. Vendors, and internal AI teams, who cannot attach their tool to revenue, margin or hours saved will be renegotiated by people who can attach it to a cheaper invoice.
The bottom line
The frontier labs spent two years competing on benchmarks. They are now competing the way enterprise software has always competed once the technology is good enough: account by account, discount by discount, in private. OpenAI is spending margin to convert Anthropic’s marquee logos before an IPO, Anthropic is defending a premium its customers are publicly questioning, and underneath both of them the open and Chinese models keep lowering the price everyone has to beat. For buyers, this is the best negotiating moment the AI market has offered. It will not stay open long, price wars end with consolidation, and the bill for today’s 90 percent discount arrives at renewal.