THE SHORT ANSWER
● Who won the account — OpenAI, with a discount of more than 30%
● Deal size — Reported at over $100 million a year
● What triggered it — Anthropic discontinuing the discount it had been giving Replit
● What it proves — At the application layer the two were substitutable enough for price to decide
Replit's chief executive Amjad Masad said on 17 October that Replit is no longer working with Anthropic and that OpenAI is now the default model. The Information reported on 24 October what it cost: more than 30% off, in an arrangement worth over $100 million a year, after Anthropic ended its own discount.
What actually changed hands
| Anthropic | OpenAI |
|---|
| Position before | Default model at Replit | Secondary |
| Position now | Not in use, per Replit's CEO | Default |
| Discount offered | Discontinued | More than 30% |
| Annual value | Was over $100M | Over $100M |
| Capability argument | Not reported as the reason | Not reported as the reason |
The interesting absence
No benchmark appears anywhere in this story. Replit builds a coding product, which is the use case where model differences are supposed to be most visible and most costly to get wrong. If the switch had been about Claude losing on code, that would have been the headline. Instead the reported trigger is a withdrawn discount.
That is a statement about the market, not about either model. It says that at the application layer - wrapped in a harness, with retries and scaffolding around it - the two were close enough that a 30% price difference outweighed whatever capability gap remained. The Shandong harness study published the same week supports the mechanism: a vendor's own harness on its own model buys no score advantage, and scaffolding choices move cost far more than accuracy.
What it means for each side
For Anthropic, the timing is awkward. It reported its first profitable month in August and filed its S-1 in October. Holding price discipline on large accounts and holding the accounts themselves are now directly in tension, and a prospectus makes both visible.
For OpenAI, buying a nine-figure account at 30% off is a customer-acquisition cost, not a margin win. It works if switching costs keep Replit there once the discount normalises.
For everyone else, the lesson is that these discounts exist and are negotiable. If a $100M account can move on price, a $100k one can ask.
The pressure that is still coming
Kimi K3.1 Pro entered the Artificial Analysis Intelligence Index at 55 against Claude Opus 5.5's 58, at $2 input and $10 output per million, with open weights promised in November. The Replit deal was a negotiation between two closed vendors. The next one may not be, and an open-weight model three points off the frontier changes what a discount has to beat.