Replit's CEO said the Anthropic work is over. A changelog still lists Claude.

A stage sentence and a changelog

TechCrunch reported that on October 16, 2026, at its Disrupt conference, Replit's chief executive Amjad Masad said the company is no longer working with Anthropic. He gave cost, partnership, and performance as the reasons. He said he thinks ChatGPT is better now, and that Replit's own benchmark shows OpenAI's Sol and Astra performing better at building applications from scratch. TechCrunch notes an OpenAI partnership announced in August, and says it had not seen an earlier public ending of the Anthropic relationship. The same article says a person familiar with the partnership, who was not named, said it was still in place, and that TechCrunch had asked Replit to clarify. EffectStory lines the stage remark up against Replit's changelog: Claude Opus 5.5 on September 25, Claude Sonnet 5.5 on October 2, and Claude Haiku 5.5 on October 16. Dealroom's feed entry repeats the stage remark and the unnamed contradiction. The artificial-intelligence desk isArtificial Intelligence.

What kind of fact a stage remark is

A CEO on a stage can describe a vendor relationship in a sentence that the company's own release notes have not caught up to, or in a sentence that overshoots the contract. This page does not know which. It knows TechCrunch printed the sentence, printed the contradiction, and said it had asked the company. Until that ask returns, the record is a conflict, not a eulogy for a partnership. The paper isFutureweb.

What a changelog entry can prove

A changelog line that adds Claude Haiku 5.5 on the same calendar day as the stage remark proves that a Replit surface still named a Claude model. It does not prove a particular contract is in force. It does make no longer working a phrase that needs a gloss: no longer the preferred vendor, no longer in new deals, or no longer in the product. Masad's sentence did not include the gloss. EffectStory's point is that the product surface still did. Software isSoftware.

Benchmarks he pointed at stay his

Sol and Astra beating other models at from-scratch applications is Masad's description of Replit's benchmark, via TechCrunch. This paper did not run it. A vendor score, even a public one, is the vendor's until a second lab can run the same check. That standard is already on this desk. Using it here keeps a conference boast from becoming a measurement. An evaluation a second lab can run isan evaluation a second lab can run.

August's OpenAI deal is context, not the ending

TechCrunch uses the August OpenAI partnership to say the new sentence was not preceded by a public Anthropic breakup. An added partnership does not logically delete another one. Many products call more than one model. That is why the changelog matters. Two vendors on a changelog is an ordinary state. A CEO saying one of them is gone is the change, and it is still a claim. Hardware isHardware.

Dealroom is a third telling of the same stage

Dealroom's item repeats TechCrunch's scene: the stage, the reasons, the benchmark sentence, the unnamed person, the September production mention. It is not an independent interview with Masad. Its value is that the conflict survived a second writeup. It does not add a document from Anthropic. This brief has no Anthropic statement. The absence is part of the record. Security is Asked directly, Who contradicted a finished split? TechCrunch says a person familiar with the partnership, unnamed, said it was still in place. Dealroom's brief of the same remarks includes that contradiction. EffectStory adds the changelog dates.Security.

What this page will not announce

This page will not say Replit has removed Claude. It will not say the partnership contract is intact. It will say Masad said the work had stopped, a person familiar told TechCrunch it had not, and a changelog still listed Claude models through October 16. Readers who need a single status have to wait for the company to answer the question TechCrunch says it asked. Networking is Asked directly, What does EffectStory say the changelog showed? It says Replit's changelog listed Claude Opus 5.5 on September 25, Claude Sonnet 5.5 on October 2, and Claude Haiku 5.5 on October 16, 2026. A model name on a changelog is not a contract, and it is hard to square with a split that already removed the vendor.Networking.

Who is in scope

Replit customers who choose a model, and anyone treating a conference keynote as a vendor obituary, are in scope. Anthropic's unstated position is in scope only as a blank. Filling the blank would be fiction. Open source is Asked directly, What reasons did Masad give? TechCrunch says he named cost, partnership, and performance. He said ChatGPT is better now, and that Replit's benchmark shows Sol and Astra doing better at building applications from scratch. That benchmark claim is his, reported by TechCrunch.Open Source.

Where the conflict sits

TechCrunch is the stage and the unnamed contradiction. EffectStory is the changelog dates. Dealroom is another telling of the TechCrunch scene. Nothing here is a contract termination notice. The short answer this page will stand on is this: Amjad Masad said on stage that Replit is no longer working with Anthropic, for cost, partnership, and performance, and he pointed at OpenAI models on Replit's benchmark. TechCrunch reported that a person familiar said the partnership was still in place. EffectStory says the changelog listed Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5, the last of those on October 16, 2026.