OpenAI Discounts Win Replit's $100M Contract and Flip Rippling's Coding Mix From Anthropic

TL;DR

The Information reported on October 24 that OpenAI is winning Anthropic's biggest coding customers with negotiated discounts below list price. Replit moved most of its AI spend to OpenAI after an offer of more than 30% off a contract worth over $100 million a year. At Rippling, CEO Parker Conrad says Anthropic's share of his engineers' coding work fell from 57% in May to 18%, while OpenAI's rose from 16% to 52%. DataCamp was offered 40% off if it also adopts Codex, and a US bank got 500 ChatGPT seats for about $40,000 a year. Anthropic booked $11.6 billion of revenue against $5.6 billion of compute in the June quarter. OpenAI projects compute spend above revenue this year.

Four accounts, one playbook

Laura Bratton and Stephanie Palazzolo's report sits behind a paywall, but the customer-level numbers are consistent across the Next Web, Decoder and Dealroom summaries, and two of the executives are on the record by name.

• Replit. A person familiar with the deal says OpenAI offered more than 30% off a contract worth over $100 million a year. Replit had planned to stay with Anthropic, which had discounted its models before, but Anthropic would not extend the discount without a bigger spending commitment.
• Rippling. In May, Anthropic models handled 57% of the coding work done by Rippling's hundreds of engineers and OpenAI models handled 16%. Conrad says those shares are now 18% and 52%, and that the company moved to cut costs.
• DataCamp. CEO Jonathan Cornelissen was offered 40% off OpenAI's models for internal development, on the condition that DataCamp also uses Codex instead of Claude Code. He told The Information the change could cut DataCamp's AI costs by 80% to 90%. That figure blends the discount with list prices that were already half of Anthropic's, so it is not a rate card.
• An unnamed US bank with more than $7 billion in assets doubled its ChatGPT seats to 500 at no extra charge. That works out to about $40,000 a year, or under $7 per user per month against a usual $20 or more.

The report also says Atomicwork is training its engineers on Codex with free credits from OpenAI and Khosla Ventures.

Anthropic's discount stops at the cap. OpenAI's does not.

The mechanism behind the switching was reported a month earlier. On September 28, The Information's Kevin McLaughlin, relayed by The Next Web, spoke with managers at software companies that buy from both labs. They put Anthropic's enterprise discounts at about 15% off list on large commitments, and said the discount ends the moment a customer burns through the tokens in its contract. Hit the cap and you either renegotiate or pay list for the overage. One executive said Anthropic's sales team calls as soon as the ceiling is reached, new deal in hand.

Picture a gym that sells you 100 visits at a discount, bills full retail from visit 101, and has the front desk phone you mid-squat with a renewal offer. OpenAI, by the same account, gives you the rest of the current month plus one more month at the old rate to agree a new contract. Amazon, Microsoft and Google usually let customers keep their discount on usage above the commitment for the rest of the term, according to Jeff Muscarella of NPI Financial.

CodeRabbit CEO Harjot Gill told the same reporter that OpenAI is now his company's main supplier, after Anthropic held that spot six months earlier; CodeRabbit spends tens of millions of dollars a year on models. Gill and another manager also said both labs have started writing "share of wallet" clauses into discount deals, which require large customers to route most of their AI budget to one platform. Anthropic did not comment for that report. OpenAI hired its first worldwide sales chief the same month, and the discount memo appears to have arrived with the desk.

List prices halved first, then the haggling started

None of this is happening against stable rate cards. On September 22, OpenAI launched GPT-6 Sol at $2 per million input tokens and $10 per million output, half of GPT-5.6 Sol's $4 and $20 promotional rate, and GPT-6 Luna at $0.10 and $0.50, per The Next Web's launch coverage. The announcement landed about 90 minutes after Anthropic shipped Claude Opus 5.5 at $4 and $20, down 20% from Opus 5, with cache reads cut from $0.50 to $0.20. Anthropic's post says it outright: "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that."

Max Woolf's timeline of the cuts, which drew more than 400 points on Hacker News on Saturday, counts drops of 20% to 80% across both labs' lineups since June and a 75% cut to Claude Fable's cache-read price. His point for coding shops: cache reads are 96% to 99% of a coding agent's tokens, so the cache line on the price sheet matters more than the headline input price. Vercel's gateway data, cited in the same post, shows cost per token at 43.3% of its April level while total spend keeps rising, because models improve faster than they get cheaper.

Negotiated discounts sit on top of those list cuts. A 30% deal on GPT-6 Sol is 30% off a price that was already halved, which is how DataCamp's CEO gets to 80% or 90% without OpenAI ever publishing a 90% discount.

The margin behind the strategy

The Information's thesis is in its headline: Anthropic's margin is OpenAI's opportunity. Anthropic's S-1, filed October 26, shows $11.6 billion of revenue in the June quarter against $5.6 billion of compute, and Fortune's read of the filing quotes the company at $2.08 of revenue per compute dollar that quarter. OpenAI projects compute spending above revenue for the year. A lab earning two dollars per compute dollar can hold its discount at 15% and rely on the cap. A lab already spending more than it makes can add another 30% off to a loss it had planned anyway.

The report's own caveats cut the other way. Anthropic still grew 45% in the third quarter, slower than in the second. It retook the lead on OpenRouter after Opus 5.5 shipped. And on Ramp's AI Index, which reads corporate card and invoice data, Anthropic's share of AI token spending rose from 53.8% in the week ending September 13 to 60.1% in the week ending October 11, while OpenAI's fell from 42.4% to 36.4%. In September, 43.8% of US businesses on Ramp paid Anthropic and 40.6% paid OpenAI. OpenAI, for its part, says its enterprise growth has accelerated.

So the data describe a handful of very large coding accounts moving on price while the broad base of paying businesses tilts the other way. Both can be true. Large accounts are where the discounts are.

Replit's version of events

Replit's CEO had already said the quiet part on stage. At TechCrunch Disrupt earlier this month, Amjad Masad said Replit was no longer working with Anthropic and that "Sol and Astra perform better on building vibe-coded applications from scratch." His summary: "It was a decision based on both cost, partnership, and performance." Over the following weekend he told TechCrunch that "Replit remains model agnostic," that users can still pick Claude, and that OpenAI "is now our preferred partner for powering Replit Agent." An insider countered that the Anthropic partnership was still in place, and Replit's own late-September blog post listed Claude Fable 5 and 5.1 in Agent's production stack.

The two statements are compatible. TechCrunch notes that most Replit customers keep the default, so switching the default moves most of the spend even if Claude stays on the menu. Masad said Replit had recently been one of Anthropic's top five customers, and that enterprise contracts are about 15% of revenue this year, up from about 2% last year.

If you buy tokens

• Ask what happens at the commit cap before you sign. The difference between list price from the first token past the cap and two months of grace is the whole story above.
• Read the share-of-wallet clause. A 40% discount that requires routing most of your budget to one vendor is lock-in with a sticker on it.
• Price the cache line, not the input line. For agentic coding, cache reads are the bill.
• Keep the routing layer. Rippling moved a majority of its coding work between vendors in five months, which only works if your tooling treats the model as a config value.
• Treat "up to 90%" as one CEO's blended estimate. The reported offers are more than 30% and 40%, on top of list prices that fell in September.

Caveats

Every number in the first section comes from a single outlet. Replit's terms are attributed to an unnamed person, and neither lab has confirmed any of the deals. Rippling's figures describe a share of one company's coding work, not spend or output quality, and Conrad's own framing is that the move was about cost. DataCamp's 80% to 90% is an estimate that mixes a discount with lower list prices and a different workload. Promotional terms expire, and the September reporting is explicit that discounts at both labs are tied to commitments that get renegotiated.

Key Takeaways

• The Information reports OpenAI won most of Replit's AI spend with a discount of more than 30% on a contract worth over $100 million a year.
• Rippling's coding work went from 57% Anthropic and 16% OpenAI in May to 18% and 52% now, per CEO Parker Conrad.
• Anthropic's enterprise discounts run about 15% off list and end at the commit cap. OpenAI gives a two-month grace window, per managers who buy from both.
• Both labs cut list prices on September 22, and negotiated discounts stack on top of those cuts.
• Anthropic made $2.08 of revenue per compute dollar in the June quarter. OpenAI projects compute above revenue this year.
• Ramp's broad data point the other way: Anthropic's token-spend share rose to 60.1% in the week ending October 11.

Sources: The Information, Anthropic's Margin Is OpenAI's Opportunity (Oct 24, 2026), The Next Web (Oct 26), The Decoder, Dealroom summary, MLQ, The Next Web on Anthropic's discount caps (Sep 28), TechCrunch, Replit makes OpenAI its default (Oct 17), The Next Web, GPT-6 Sol and Luna pricing (Sep 22), Anthropic, Claude Opus 5.5, Max Woolf, cheap LLMs, Hacker News thread, Anthropic Form S-1, Fortune on the S-1, Ramp AI Index