Three kinds of number, only two of them granted

The biggest discount in this weekend's OpenAI pricing story is not a discount anyone has been shown receiving. The Information reported that DataCamp obtained a 40% reduction for models used in internal product development, conditional on also using Codex. CEO Jonathan Cornelissen estimated that costs could fall by 80–90% if comparable terms were extended to models inside the products DataCamp sells. He was seeking that extension. The Decoder's October 25 headline turned the conditional estimate into an apparent OpenAI offer of up to 90% off. The reporting available to ALFW does not establish such an offer.

Read the distinction precisely, because it is the whole story. The 80–90% figure is an estimated reduction in DataCamp's costs under a hypothetical extension of terms it already has on a narrower use. It is not an estimated discount rate in a signed contract, and it is not a rate OpenAI has published for anyone. What has actually been granted, on the reporting, is smaller and more specific: a 40% reduction fenced to internal product development and tied to adopting a second product, Codex.

What has been reported at Replit and Rippling

According to The Information, as reported by subsequent outlets, Replit shifted most of its AI spending to OpenAI under an annual contract reportedly worth more than $100 million, with a discount exceeding 30%. The reported sequence matters more than the headline percentage. Anthropic had declined to extend the discount it was giving Replit unless Replit committed to higher spending; OpenAI's offer arrived after that. The contract figure is reported, not announced: neither company has published terms, and the reporting does not establish whether the $100 million is measured before or after the discount, so no savings figure can be calculated from it.

At Rippling, chief executive Parker Conrad said OpenAI's share of his company's coding workload rose from 16% in May to 52%, while Anthropic's fell from 57% to 18%. The two named providers account for 73% of the earlier reported coding workload and 70% of the later one. The remaining 27% and 30% are not allocated between providers in the cited figures. This is a shift in the largest slice of one company's coding workload, reported by its chief executive — not a two-vendor seesaw, and not market-wide data.

The bank item shows the third kind of number. The Decoder also relayed a report about an unnamed US bank whose ChatGPT seats were doubled to 500 without additional charge, bringing the reported effective cost below $7 per user. The institution, agreement and billing terms have not been publicly disclosed. The Decoder's own report puts the usual price at $20 or more per user. A seat expansion on unpublished terms is a real concession if the reporting is right, but it is not a rate card, and no further discount percentage can be derived from it.

The breakup that was a default change

Replit is also where this story can mislead, because an earlier, louder version of it was wrong in a specific way. On October 16, at TechCrunch Disrupt, Replit chief executive Amjad Masad said the company was no longer working with Anthropic, citing cost, partnership and performance, and added, "I think ChatGPT is better now." He also said Replit had recently been one of Anthropic's top five customers. Within days he had qualified the claim: Replit "remains model agnostic," OpenAI is now its preferred partner for powering Replit Agent, and users can still choose Claude. A person familiar with the partnership told TechCrunch it was still in place, and TechCrunch softened its own headline from a breakup to a default change.

The documentation settles what can be settled. Replit's model selector documentation, checked on October 25, lists seven distinct Claude models available to users now: Claude Sonnet 5 and Claude Sonnet 4.6 in Power Mode, and Claude Fable 5.1, Claude Opus 5.5, Claude Opus 4.8, Claude Opus 5 and Claude Opus 5 Fast in Max Mode, alongside OpenAI's models and Kimi K3. Tables circulating this weekend that mark Anthropic as "not in use" at Replit are contradicted by the vendor's own documentation. What is established publicly is a change in Replit Agent's preferred provider and default, alongside The Information's reporting of a substantial spending shift. That does not mean Claude has disappeared from Replit. Defaults are not a detail — most users never change one — so the shift is a commercial loss for Anthropic without being an exit.

Why the discounts exist, on the reporting

The Decoder's summary adds the context that keeps this from reading as pure margin destruction: OpenAI has cut some of its own serving costs with its newer Sol and Luna model lines, so part of the discounting is efficiency being passed through, and part, on The Decoder's account, is share being bought at the expense of margins. An investor memo cited in the same reporting says Anthropic's growth slowed between the second and third quarters. Both points are secondary-reported context, not audited figures; neither establishes that the discounts are sustainable or that Anthropic is financially impaired. None of it tells a buyer what they will pay.

What a reader should take from it

If your company buys model access at any scale, the actionable fact in this reporting is not 90%. It is that a customer spending at Replit's level lost its discount by standing still, and recovered a larger one by being willing to move its default. Discounts in this market appear to expire, to be fenced by use case, and to be conditional on adopting adjacent products, as DataCamp's Codex condition shows. Each of those qualifiers is worth more in a negotiation than the headline percentage, and each one is exactly what a headline percentage strips away.

Sources

  1. [1] The Information, “Anthropic's Margin Is OpenAI's Opportunity” (reported October 24, 2026; paywalled — figures as relayed by Techmeme, AI Weekly, The Decoder and Sina Finance) — the Replit contract and discount, the Anthropic trigger, the Rippling coding-workload shares and the DataCamp termsRead source
  2. [2] The Decoder, “OpenAI offers discounts up to 90 percent to poach Anthropic's biggest customers” (October 25, 2026) — the 90% headline under examination; its report attributes an 80 to 90% cost reduction to DataCamp's CEO without the conditional-extension account carried by the fuller relays; also the source of the bank seat item and the Sol and Luna cost contextRead source
  3. [3] Sina Finance reporting via Binance Square, “OpenAI Discounts Win Coding Work From Anthropic Customers” (October 25, 2026) — the fullest relay of the DataCamp distinction: 40% granted on internal product-development models provided Codex is also used; the extension to product models being sought; the 80 to 90% figure stated as Cornelissen's estimate if the offer appliedRead source
  4. [4] AI Weekly, “Replit Shifts $100M+ AI Spend to OpenAI After 30% Price Cut” (published 2026-10-24T20:01:48Z) — the discount trigger, the Masad walk-back and the default-stickiness framingRead source
  5. [5] Replit documentation, “Model selector” (official; read directly by ALFW and checked October 25, 2026 — seven Claude models listed across Power and Max modes)Read source
  6. [6] TechCrunch, Replit / Masad at Disrupt (October 16–17, 2026) — “no longer working with Anthropic,” the model-agnostic clarification, the person-familiar dispute and the headline revision to a default change, as relayed by AI Weekly and other summariesRead source