Executive Summary

This article reads the move by the AI coding platform Replit, which changed its default model from Anthropic's Claude to OpenAI, not as news about a supplier changing hands but as a case where one decision carries two separate explanations. Co-founder and chief executive Amjad Masad named three reasons on stage on October 16: cost, partnership and performance. Over the weekend he revised the remark, saying Replit is still tied to no particular model, and adding that what changed is the default inside Replit Agent and that users can still pick Claude themselves.

The Information then reported that OpenAI won a contract worth more than $100 million a year with a discount of over 30 percent, and that the opening appeared after Anthropic declined to extend an existing discount. The reasons the company gave aloud and the terms the reporting set down do not contradict each other. The performance may well be better and the price may well have been cut. What cannot be settled from outside is which of the two counted for more, and neither Replit nor OpenAI nor Anthropic has ever confirmed the figures. The performance side sits in a similar position. The public benchmark Masad pointed to does exist, but the scores published on it predate the models he named, and an Anthropic model sat at the top.

Sections 1 through 4 follow what can be verified in the TechCrunch report, Masad's public remarks, the coverage citing The Information, and the benchmark material Replit has published. The judgment in section 4 that the two accounts cannot finally be placed side by side, and the question in section 5 about what should be written down when a default changes, are this article's own interpretation and appear in none of those sources.

Key Figures

Every number attached to this switch came from one side only. The last card is about that.

Source: Techmeme, summarizing The Information · The Decoder · TechCrunch

Over 30%

Discount OpenAI is reported to have offered Replit

The figure The Information attributed to a source it did not name

$100M+ a year

Annual size of the contract that discount was attached to

No contract has been made public, and readings differ on whether these are offered or signed terms

7

Claude models that reporting counted in Replit's documentation

They did not come off the list; they came off the default

Zero

Companies that have confirmed the discount terms

Replit, OpenAI and Anthropic have neither affirmed nor denied the figures

1

It Wasn't a Breakup, It Was a Default Swap

On the TechCrunch Disrupt stage on October 16, Masad said Replit is no longer working with Anthropic. Asked why, he answered: "It was a decision based on both cost, partnership, and performance." Weighing all of it together, he added, OpenAI felt like the better partner for now. On performance he was blunt: "I think ChatGPT is better now." His grounds were not only an internal evaluation. Replit publishes a public benchmark as well, he said, and on it OpenAI's Sol and Astra come out consistently better at building applications from scratch.

The rebuttal ran inside the same article. A person familiar with the matter, speaking on condition of anonymity, said the partnership remains in place, and the piece noted that Replit's homepage, blog and social accounts had been promoting Anthropic models as recently as a few weeks earlier. A blog post from late September stated that Anthropic models were running in Replit Agent's production environment.

Over the weekend Masad revised the remark. Replit remains model agnostic, he said; users can still choose Anthropic's Claude and other supported models, and OpenAI has become the preferred partner currently powering Replit Agent on the strength of performance, economics and partnership. All that moved is which model Replit Agent reaches for when nothing is configured. The seven Claude models that reporting counted in Replit's documentation are still on the list, and using one now means opening the menu and picking it.

The remarks were not the only thing that moved. TechCrunch appended a note to the article saying that the piece, first published at 2:22 p.m. Pacific on October 16, had been updated to add context on Replit's partnerships with Anthropic and OpenAI, and that its title had been updated as well. The headline now reads that Replit makes OpenAI its default over Anthropic. What the first headline said still survives in the address: the slug at the end of the URL reads that Replit is no longer working with Anthropic.

Over three days one sentence narrowed from a breakup to a change of default. At each step the scope was redrawn rather than new facts added, and across the whole of it not a single number was made public.

2

Eight Days Later, a Different Number Arrives

Between October 24 and 25, The Information brought a different account. Replit had shifted most of its AI spending from Anthropic to OpenAI, the report said, and behind that shift lay a discount of more than 30 percent and a contract worth over $100 million a year. The headline framed Anthropic's margin as OpenAI's opportunity. The reporting rested on one unnamed source, and it placed one event earlier in the sequence: Anthropic had not extended Replit's existing discount, reportedly declining to renew it without a higher spending commitment.

These figures still have one foot off the ground. The remarks on stage survive on video and in print, but the contract terms are a single outlet's exclusive, routed through an unnamed source, and none of the three companies has affirmed or denied them. One explanatory outlet that went looking found nothing in OpenAI's or Anthropic's own announcements that confirmed the specific terms. Alongside that came the point that it is unclear whether the numbers describe an offer or a signed price. Read either way, 30 percent of $100 million is more than $30 million.

One more timing overlaps here. Anthropic is reported to be pursuing an initial public offering in mid-November, with the roadshow understood to begin in the second week of the month. In the same week, the engineer leading Claude Code and the developer who built the desktop app announced their departures a day apart. Nothing among these events has been shown to cause anything else. What can be verified is the overlap of dates.

3

Replit and Rippling Are Two Different Companies

Two other companies appear in the same reporting, and this is where the numbers get mixed up most easily. The figures showing a share of coding work flipping from 16 percent to 52 percent belong to Rippling, not Replit. Rippling is a separate company that builds HR and IT management software, and the numbers came from its chief executive, Parker Conrad, directly. Over the same period Anthropic's share fell from 57 percent to 18 percent. Cost was the reason he gave. For Replit, the only figures made public are the discount and the size of the contract; no before-and-after usage share for Replit appears in the reporting at all.

DataCamp also carries two numbers. The contract term the reporting describes is a 40 percent discount received in exchange for using OpenAI's Codex first, while the 80 to 90 percent its chief executive, Jonathan Cornelissen, spoke of is the cost reduction he experienced. The first is a discount rate in a contract and the second is a change in money actually spent, so they are not figures on the same axis. The 90 percent that turned up in one outlet's headline comes from the upper end of that second number.

Putting all three companies' figures in one place makes clear which number belongs to which company, and what it measures.

Company Reported figures What those figures measure
Replit Discount over 30%, more than $100M a year Price terms of a contract. No usage share has ever been made public
Rippling 16%→52%, 57%→18% Share of coding work each vendor's models handled. Stated by the CEO himself
DataCamp 40% discount / 80–90% savings The first is a contract discount rate, the second a cost change the CEO described

A signal running the other way landed in the same month. The index Ramp published on October 14 recorded that, as of September, 43.8 percent of businesses were paying Anthropic against 40.6 percent paying OpenAI. Anthropic rose 0.25 of a percentage point while OpenAI rose 1.12, so the gap narrowed, but the order held. These are shares of businesses that paid, not market share, and because many pay both vendors the two add up to more than 100. It is a number to read with its axis in mind. In the same index, Anthropic's share of token spending for the week ending October 11 was 60.1 percent, a far wider lead than the one it holds by number of businesses.

4

The Two Accounts Cannot Be Weighed From Outside

The explanation that performance improved and the report that the price was cut do not push each other out. If an option was both cheaper and better, both are true. The difficulty comes after that. Working out which of the three drove the decision means putting the evidence for each side in one place. Masad named all three at once, yet the headlines carrying his remark mostly picked one. Some picked performance and some picked the discount. The headlines narrowed first, before the remark did.

Start with performance. The grounds he gave went beyond an internal evaluation, and the benchmark he invoked on stage is real: "We publish a public benchmark as well, and consistently, Sol and Astra perform better on building vibe-coded applications from scratch." ViBench, built by Replit with the Georgian AI Lab and researchers at Carnegie Mellon University, hands models requirement documents drawn from real usage logs and then runs the resulting apps automatically to see whether they met the requirements. The code is open and a paper is out. Hiding the method is not a charge this decision can be criticized on.

The empty cell is not the method but the comparison he cited. The scores in the ViBench paper come from tasks drawn across 15 apps, run over nine models, with 105 resulting outputs graded, and first place went to Anthropic's Opus 4.6. Its first-attempt pass rate was 46 percent, with OpenAI's GPT-5.2 behind it at 42 percent. That paper predates Sol and Astra, so it does not refute what Masad said. His remark did not say which version of the result he meant, or how many runs produced it, and the public material available as this article is written does not hold the same comparison either.

Source: Replit, introducing ViBench · ViBench paper (ACM CAIS 2026)

The price side stands differently. Here it is one outlet relaying the account of someone who says they saw the terms, and the contract itself is not public. One side points at a testing ground that is open while the score it cites cannot be found on that ground; the other has no document to check against at all. There is still no shared material on which to put the two on one scale.

One decision, two explanations Both can be true, but there is no shared record to weigh them on What the company said Cost · partnership · performance Grounded in a public benchmark Cited score not found there What reporting recorded Over 30% off · $100M+ a year Grounded in an unnamed source Contract is not public cannot compare What can be checked from outside The default changed, and the outputs change with it
▲ The two accounts attached to one decision and how public each of them is — a Pebblous diagram of the reporting and the remarks on stage

So the facts an outsider can hold in this episode come down to one. The default changed, and for most users, the ones who never open the menu, that single change is the whole of it. Why it changed stays between the two accounts. What changed shows up in the user's output, and that second one is the kind of fact that can be checked.

5

Why Pebblous Is Watching This Switch

A price fight between suppliers is not something for Pebblous to follow. The reason to look is the place where the swap happened. What changed is one line of configuration, and what changed as a result is all the code built with that configuration. A person who put a request into Replit Agent yesterday and a person who puts the same request in today receive answers from different models. Neither of them knows what they got, and nothing in the result they received says so.

The reason this asymmetry will grow is inside this reporting. The cheaper it gets to swap suppliers, the more often companies swap them. Rippling flipped the share of its coding work in five months, and DataCamp cut its price in exchange for reaching for a particular tool first. As the interval between swaps shortens, outputs built inside the same product come from different models depending on when they were made. When quality drops and there is no material to separate the model from every other change, finding the cause falls back to guesswork.

If you are putting AI-generated output into a product, four things are worth checking.

  • Does the output carry which model and which version made it? The vendor's name is not enough. Within the same company, a change of version changes the result
  • Is there a record of when the default changed? Without the date of the change, a wobble in a quality metric cannot be lined up against it
  • Can you tell apart the cases where a user chose the model from the cases where the default was left alone? Mixing the two removes the comparison group
  • Can the same task be run again across the swap? If the inputs to re-run are gone, neither the claim that performance improved nor the claim that it got worse can be checked

None of the four says to stop changing models. They say to change them and write down what changed in your own data. Whether Replit received a discount cannot be verified from outside, but which model produced your own company's output is sitting somewhere you can verify it. In an environment where suppliers do not publish their reasons, the record you can trust comes from that side.

Thank you for reading this far. The facts in this article were checked by comparing the sources listed in the references below against one another, and items the companies have not confirmed, such as the discount rate and the contract size, are written down as what the reporting says. If you already keep a model history alongside AI output, we would like to hear which fields you record.

Pebblous Data Communication Team
October 26, 2026

R

References

Reporting

Primary Sources & Academic Paper