Weekend Edition · Sat 17 Oct 2026
Replit’s CEO Says the Anthropic Partnership Is Over. A Conflicting Account Says Otherwise
Replit chief executive Amjad Masad said on stage at TechCrunch Disrupt that the coding platform is no longer working with Anthropic, citing cost, partnership and performance. Replit's own documentation still listed seven Claude models among the 11 that paying users can select when the Journal checked it on Saturday, leaving the status of the relationship unresolved.
Chanan Zevin - Chief Editor and Head of Desks
Published
What the chief executive said
Amjad Masad, chief executive and co-founder of Replit, said on Thursday, 15 October, that his company is no longer working with Anthropic. He made the remark on stage at TechCrunch Disrupt, according to a TechCrunch report by Lucas Ropek published on 16 October. Masad said the decision rested on "cost, partnership, and performance." He added: "I think ChatGPT is better now." He said OpenAI's Sol and Astra models consistently performed better on Replit's internal benchmark and on a public benchmark the company publishes for building applications from scratch.
An account disputed at once
The claim was contested in the same report. TechCrunch said insiders familiar with the partnership rejected Masad's comments, and one person who asked not to be named said the partnership was still in place. The Journal found no public statement from Replit or Anthropic clarifying the contractual position before its research cutoff at 12:30 GMT+3 on Saturday. This article therefore treats the end of the partnership as the chief executive's claim, not as an established fact.
Why Claude mattered to Replit
The relationship has been important to Replit's growth. Under its partnership with Anthropic, Replit used the AI lab's Claude models to run its platform, TechCrunch reported, and that partnership was said to have helped the company grow its revenue significantly. AInformation reported that on 17 June Replit also became callable from inside Claude itself. A break would therefore touch both the engine behind Replit Agent and a distribution channel that brought Claude users to Replit.
Partnership versus product
A commercial partnership and the models available inside a product are different things. A company can end a strategic arrangement, such as joint marketing, preferred pricing or co-development, while still buying model access through a standard contract. It can also keep a partnership in place while shifting most of its traffic to another supplier. Masad's remarks, as reported, do not say which of these changed. That gap is the core of the story, and it is why a single sentence on a conference stage has produced two incompatible readings.
What the documentation shows
The Journal checked Replit's public documentation directly. At 12:20 GMT+3 on 17 October, the Model selector page at docs.replit.com listed 11 models that paying users can choose in Replit Agent's Power Mode and Max Mode. Seven of the 11 were Anthropic Claude models. Three were OpenAI GPT-6 models, and one was Kimi K3 from Moonshot AI. Chart 1 shows the split by developer. This is a verified finding about what the documentation says, and only about that.
Which Claude models are listed
In Power Mode, the page listed Claude Sonnet 5 and Claude Sonnet 4.6 alongside GPT-6 Sol and GPT-6 Luna Fast. In Max Mode, it listed Claude Fable 5.1, Claude Opus 5.5, Claude Opus 4.8, Claude Opus 5 and Claude Opus 5 Fast, alongside GPT-6 Astra and Kimi K3. The page says the Fast variants require a Pro or Enterprise plan. Japanese technology outlet AInformation reported the same count of seven Claude models out of 11 after checking the page on 17 October.
Who can choose a model
Manual model choice is limited to paying users. Replit's documentation says builders on the Core and Pro plans can choose an available model in Power Mode or Max Mode, while Free Mode always uses automatic routing. AInformation reported that the free Starter tier cannot use Power Mode or Max Mode at all. In practice, then, only paying customers can deliberately select a Claude model, and they are the users who would notice first if the list changed.
Limits of that evidence
The documentation does not prove what any individual user sees. The page itself says model selection is still rolling out and that availability can vary by rollout, organisation settings and authorisation. It calls the list inside each user's own Model selector "the source of truth" for that account. The Journal did not test model availability inside a paid Replit workspace. A documentation page can lag behind a product change, and it can also run ahead of one.
Earlier public signals
TechCrunch reported that a Replit blog post in late September cited Claude Fable 5 and Claude Fable 5.1 as models running in Replit Agent's production environment. It also reported that a Replit product employee promoted Anthropic's flagship Opus 5.5 model on X around the same time. Those references support the view that Claude was in active use only weeks ago. They do not establish the contractual position today, and they do not show how much of Replit's traffic Claude now serves.
A turn towards OpenAI
There is public evidence of a shift towards OpenAI. Replit announced a partnership with OpenAI in August, according to TechCrunch. AInformation reported that Replit's Free Mode, launched on 18 August, was branded "With OpenAI," and that a 29 September Replit post described GPT-6 Astra handing smaller tasks to cheaper sub-agents. Free Mode does not let users pick a model. Replit routes those requests automatically, so the mix of models serving free users cannot be read from the documentation. Table 1 sets out the dated public record.
| Date | Event | Basis |
|---|---|---|
| 17 Jun 2026 | Replit becomes callable from inside Claude | AInformation, 17 Oct (attributed) |
| Aug 2026 | Replit announces partnership with OpenAI | TechCrunch, 16 Oct |
| 18 Aug 2026 | Free Mode launched, branded "With OpenAI" | AInformation, 17 Oct (attributed) |
| Late Sep 2026 | Replit blog cites Claude Fable 5 and 5.1 in Agent production | TechCrunch, 16 Oct |
| 29 Sep 2026 | Replit post on GPT-6 Astra delegating to sub-agents | AInformation, 17 Oct (attributed) |
| 15 Oct 2026 | Masad: Replit "no longer working with Anthropic" | TechCrunch, 16 Oct |
| 16 Oct 2026 | Unnamed insider: partnership still in place | TechCrunch, 16 Oct |
| 17 Oct 2026 | Docs list 7 Claude models among 11 selectable | Journal check, 12:20 GMT+3 |
Cost as a motive
Masad named cost first. Replit sells a browser-based platform on which users build and deploy apps, and in the Journal's analysis model inference is a large input cost for any agentic coding tool. The Journal has not seen Replit's supplier pricing. Any estimate of what a switch would save is therefore analysis, not reported fact. What is documented is the price users pay: AInformation, citing Replit's pricing page on 17 October, reported that the Core plan costs $20 a month and the Pro plan $100 a month.
The benchmark claim
Masad's performance argument rests on benchmarks that the Journal has not reviewed. He said Replit uses an internal benchmark to assess metrics such as performance and also publishes a public one, on which he said Sol and Astra "consistently" performed better at building applications from scratch. Benchmark results depend heavily on the task set, the prompting and the date of the test. Until the figures are examined, the claim that OpenAI's models are better for Replit's users is Masad's assessment.
Enterprise growth context
Masad also gave growth figures on stage. He said enterprise contracts were about 2% of revenue last year and about 15% this year, and that more than 50% of revenue now comes from business customers. He said Replit is not yet sharing annual recurring revenue figures. Enterprise buyers often require predictable model choices written into contracts. In the Journal's analysis, that raises the stakes of any ambiguity about which models will remain available, and for how long.
The habitat argument
Masad used the same appearance to argue that agents need what he called a "habitat." "That habitat has good orchestration, good error recovery, good safety and security and sandboxing, good networking, good identity and authorization," he said, according to TechCrunch. The remark presents Replit's infrastructure, rather than any single model, as its product. It also echoes the containment guidance for AI agents that US security agencies published on Friday, which the Journal examines in today's editorial.
What users can check
For developers, the practical step is to check the Model selector in their own workspace rather than rely on headlines. Teams that have written a specific Claude model into build instructions may want to confirm that an alternative model produces acceptable results before any change arrives. AInformation reported that Replit's documentation describes a beta Compare models feature that sends the same prompt to up to four models. It also noted that each comparison consumes credits for every model used.
What would settle it
Three things would resolve the conflict. A statement from Replit or Anthropic on the status of their commercial agreement would answer the partnership question directly. A dated change to the Model selector documentation would show whether Claude models are being withdrawn. Release notes or a deprecation notice would show when, and for which plans. None of these had appeared at the Journal's research cutoff. Until one does, both readings remain possible.
What to watch
Watch Replit's documentation and changelog for changes to the model list. Watch for any written clarification from Masad, who made the remark in a live interview rather than in a formal announcement. Watch whether Anthropic comments on its customer relationships with coding platforms. And watch whether other multi-model coding tools follow with similar public shifts. For now, the story is a contradiction between a chief executive's words and his own company's product documentation, and the documentation is the part that can be checked.
Sources
- TechCrunch, Lucas Ropek, "Replit is no longer working with Anthropic, CEO says", 16 Oct 2026, 2:22 p.m. PDT · techcrunch.com
- Replit documentation, Model selector, checked by the Journal 17 Oct 2026 12:20 GMT+3 · docs.replit.com
- AInformation, "Replit、Anthropic離れ表明 Claudeは11モデル中7つ残る", 17 Oct 2026 · ainformation.jp