Replit's CEO says the platform is no longer working with Anthropic — while Replit's own model-selector docs still list Claude options, and TechCrunch records insiders disputing the split

2 sources

On October 15, 2026 (Thursday, calendar-verified) at TechCrunch Disrupt, Replit's CEO and co-founder Amjad Masad said Replit is no longer working with Anthropic. Per TechCrunch (Lucas Ropek, 2:22 PM PDT October 16): "Replit, the AI coding platform that has become a favorite of startup founders and developers, is no longer working with Anthropic, the company's CEO and co-founder Amjad Masad said Thursday while onstage at TechCrunch Disrupt." His case: "I think ChatGPT is better now," Masad said — and: "We publish a public benchmark as well, and consistently, Sol and Astra perform better on building vibe-coded applications from scratch. It was a decision based on both cost, partnership, and performance. And for all these reasons, we just felt like they were a better partner for now." Per Replit's own model-selector docs (fetched October 18; the page's dateModified reads 2026-09-25), "GPT-6 Sol" is listed in Power Mode and "GPT-6 Astra" in Max Mode — and the same table still lists "Claude Sonnet 5", "Claude Sonnet 4.6", "Claude Fable 5.1", "Claude Opus 5.5", "Claude Opus 4.8", "Claude Opus 5", and "Claude Opus 5 Fast" alongside "GPT-6 Luna Fast" and "Kimi K3". Per TechCrunch, the pushback is on the record: "Masad's comments were rejected by insiders familiar with the partnership", one person familiar — who asked not to be named — said the partnership was still in place, and the evidence paragraph cites a late-September blog post naming Claude Fable 5 and Claude Fable 5.1 in Replit Agent's production environment plus a product employee's X post about Opus 5.5. Per TechCrunch, Replit announced a partnership with OpenAI in August, and "That partnership was said to have helped the company grow its revenue significantly" — the Anthropic one. Masad on stage: "We're not sharing ARR numbers just yet, but in terms of growth, like, I would say, enterprise was like 2% of revenue last year," and "This year, I would say like 15% of our revenue is enterprise contracts, and more than 50% of our revenue is B2B."

Why now

The coding-agent platform your audience builds on just turned into a dependency story — and the receipts are public. Per TechCrunch, Masad said onstage Thursday (October 15): "I think ChatGPT is better now," — and the claim landed with the dispute attached: "Masad's comments were rejected by insiders familiar with the partnership", one unnamed person said the partnership was still in place. Second layer, the evidence is checkable by anyone: this run opened Replit's model-selector docs on October 18 and the table still lists seven Claude options plus Kimi K3 across Power and Max modes — including "Claude Fable 5.1" and "Claude Opus 5.5" — alongside "GPT-6 Sol" and "GPT-6 Astra". Third layer, the business frame is the CEO's own: "This year, I would say like 15% of our revenue is enterprise contracts, and more than 50% of our revenue is B2B." Fourth layer, the platform lesson is Masad's own metaphor: "agents need a habitat for them to exist and do good work," — per his quote, that habitat has "good orchestration, good error recovery, good safety and security and sandboxing, good networking, good identity and authorization". Per TechCrunch, the story is still moving: "TechCrunch reached back out to Replit for clarification and will update this story if we hear back."

Why it is worth publishing

The most platform-relevant card of the day for anyone who builds on coding agents: the model brains under a vibe-coding platform are a business decision that can flip with one onstage answer, and the honest story is the gap between the claim and the docs. The differentiation play is the two-layer report: what the CEO said (with the outlet's dispute on the record) beside what Replit's own documentation lists today (Claude still selectable in Power and Max modes). Creators who publish the claim-plus-receipts version own the dependency beat; everyone else will write 'Replit dumps Anthropic', which no opened source supports as settled fact.

Evidence basis

Two opened sources: TechCrunch (Lucas Ropek, byline 2:22 PM PDT October 16, Friday calendar-verified; article:published_time meta 2026-10-16T21:22:28+00:00 with JSON-LD datePublished matching; opened and read in full via raw HTML) and Replit's model-selector docs (fetched October 18; JSON-LD dateModified 2026-09-25T09:31:53.441Z; table verified verbatim in the raw HTML). Calendar checks: October 15, 2026 = Thursday (the onstage anchor); October 16 = Friday (article publish); October 18 (this card's publish date) = Sunday. Attribution discipline: the split is Masad's onstage claim, not a settled fact — TechCrunch's own story records the insider dispute and is still seeking clarification; the card never writes 'Replit has ended its Anthropic partnership' as fact. Anthropic is silent in every opened source — no statement, no response, no confirmation. The Sol-and-Astra identification as GPT-6 models rides Replit's docs table (the article never names their maker); the docs table verifies selectable models as listed on a page stamped September 25 — never current serving traffic. The revenue figures are Masad's own on-stage characterizations; the growth sentence is TechCrunch's passive 'was said to have helped ... grow its revenue significantly' — no multiplier appears in any opened source (a 10x figure lives in a linked Inc. article this run did not open, so none is written). replit.com and its blog posts were unreachable from this run's network (TLS reset), so the late-September blog-post and X-post facts ride TechCrunch's evidence paragraph (09-18 unopened-origin precedent). The web-search quota was exhausted this run — sources were reached by fetching listing pages and articles directly (09-08 fallback precedent).

“Replit's CEO says it's done with Anthropic — but I opened Replit's own docs today, and Claude is still right there in the model list.”

Angle

Frame it as 'your coding platform's brains are a business decision — here's the claim, the dispute, and the check-it-yourself receipt' in three beats. Beat one, the claim: per TechCrunch, Masad said onstage at Disrupt that Replit is no longer working with Anthropic, because — per his quote — "consistently, Sol and Astra perform better on building vibe-coded applications from scratch" and "It was a decision based on both cost, partnership, and performance." Beat two, the dispute: per TechCrunch, insiders rejected the comments and one unnamed person said the partnership was still in place. Beat three, the receipt anyone can open: Replit's model-selector docs still list "Claude Sonnet 5", "Claude Sonnet 4.6", "Claude Fable 5.1", "Claude Opus 5.5", "Claude Opus 4.8", "Claude Opus 5", and "Claude Opus 5 Fast" across Power and Max modes, alongside "GPT-6 Sol" and "GPT-6 Astra" — say on camera that the page's own stamp is September 25, so it shows the lineup, not today's traffic.

Format

Short talking-head video

Demo idea

A claim-versus-docs walkthrough: play or read Masad's onstage line, then open docs.replit.com/features/agent/model-selector live and scroll the table — GPT-6 Sol and GPT-6 Astra beside Claude Sonnet 5, Claude Fable 5.1, Claude Opus 5.5 and the rest — ending on the two caveats: the page's dateModified stamp reads September 25, and TechCrunch's story records insiders disputing the split.

Platform notes

The split stays a CEO claim everywhere — TechCrunch's own story records insiders rejecting it and one unnamed person saying the partnership is still in place; never 'Replit has ended its Anthropic partnership' as fact, never 'Anthropic confirmed' anything (no Anthropic statement exists in any opened source). The docs table stays 'listed, page stamped September 25' — never 'still running on Claude', never today's serving lineup. Sol and Astra stay docs-identified GPT-6 models (the article never names their maker), and Masad's benchmark comparison stays his own characterization. The revenue figures — 2% to 15% enterprise, more than 50% B2B — stay Masad's own on-stage numbers with ARR unshared; no multiplier anywhere (the 10x figure lives in a linked Inc. article this run did not open). The model lineup churns — the page's own screenshots show older names. No security vocabulary enters; this is a business story. zh scripts keep Replit, Anthropic, OpenAI, Claude, GPT-6, Sol, Astra, Kimi in English, keep 「CEO 称」 framing on every split sentence, and never write 「已与 Anthropic 分手」 as fact or 「Anthropic 回应」 anything.

Usable claims

On October 15, 2026 (Thursday, calendar-verified) at TechCrunch Disrupt, Replit's CEO and co-founder Amjad Masad said Replit is no longer working with Anthropic. Per TechCrunch (Lucas Ropek, 2:22 PM PDT October 16): "Replit, the AI coding platform that has become a favorite of startup founders and developers, is no longer working with Anthropic, the company's CEO and co-founder Amjad Masad said Thursday while onstage at TechCrunch Disrupt." Masad's model-quality case: "I think ChatGPT is better now," Masad said — and: "We publish a public benchmark as well, and consistently, Sol and Astra perform better on building vibe-coded applications from scratch. It was a decision based on both cost, partnership, and performance. And for all these reasons, we just felt like they were a better partner for now." Per Replit's own model-selector documentation (fetched October 18; the page's dateModified reads 2026-09-25), "GPT-6 Sol" is listed in Power Mode and "GPT-6 Astra" in Max Mode — while the same table still lists "Claude Sonnet 5" and "Claude Sonnet 4.6" in Power Mode and "Claude Fable 5.1", "Claude Opus 5.5", "Claude Opus 4.8", "Claude Opus 5", and "Claude Opus 5 Fast" in Max Mode, alongside "GPT-6 Luna Fast" and "Kimi K3". The dispute, per TechCrunch: Masad's comments were rejected by insiders familiar with the partnership, one person familiar — who asked not to be named — said the partnership was still in place, and the outlet's evidence paragraph cites a late-September blog post naming Anthropic's Claude Fable 5 and Claude Fable 5.1 as models running in Replit Agent's production environment plus a product employee's X post about Opus 5.5. Per TechCrunch, Replit announced a partnership with OpenAI in August, and under its Anthropic partnership Replit used the AI lab's Claude models to run its platform — "That partnership was said to have helped the company grow its revenue significantly." Masad on stage: "We're not sharing ARR numbers just yet, but in terms of growth, like, I would say, enterprise was like 2% of revenue last year," and "This year, I would say like 15% of our revenue is enterprise contracts, and more than 50% of our revenue is B2B." Per TechCrunch, it "reached back out to Replit for clarification and will update this story if we hear back".

Evidence pipeline

From the news

Breakdown

A vendor-dependency story where claim-versus-receipts discipline is the content. Layer one, the claim frame: "no longer working with Anthropic" is Masad's onstage sentence at Disrupt — and TechCrunch's own story attaches the pushback: "Masad's comments were rejected by insiders familiar with the partnership", one unnamed person said the partnership was still in place, and the outlet is still seeking clarification. The card writes the split only inside the Masad-said frame. Layer two, the receipt is real but bounded: this run opened Replit's model-selector docs on October 18 — "GPT-6 Sol" in Power Mode, "GPT-6 Astra" in Max Mode, and the same table still listing "Claude Sonnet 5", "Claude Sonnet 4.6", "Claude Fable 5.1", "Claude Opus 5.5", "Claude Opus 4.8", "Claude Opus 5", and "Claude Opus 5 Fast" — but the page's own dateModified reads 2026-09-25, so the table proves the lineup as listed, never current serving traffic. Layer three, the maker discipline: the article never says whose models Sol and Astra are; the GPT-6 identification rides Replit's own docs table, and Masad's benchmark comparison stays his own characterization. Layer four, the silence rule: no Anthropic statement exists in any opened source — Anthropic is never written as confirming, denying, or responding. Layer five, the numbers: 2% to 15% enterprise and more-than-50% B2B are Masad's own on-stage figures with ARR unshared, and the growth sentence is TechCrunch's passive "was said to have helped the company grow its revenue significantly" — no multiplier, because no opened source carries one (the 10x lives in a linked Inc. article this run did not open). Editor's rules: a disputed claim stays disputed, a lineup stays a lineup, an executive's numbers stay executive-labeled, and silence stays silence.

Sources

Risks

Before publishing, re-check each layer: every split sentence carries the Masad-said frame with the dispute attached; Anthropic stays unquoted and uncharacterized; the docs table stays 'listed, page stamped September 25' — never 'still running'; Sol and Astra stay docs-identified GPT-6 models with the benchmark comparison vendor-labeled; the revenue figures stay Masad's own characterizations with no multiplier; the lineup stays churn-prone; and no security vocabulary enters. If your script compresses any of these, cut the detail rather than round it.

Demo ideas

Claim-versus-docs walkthrough: Masad's onstage line beside the live model-selector table, closing on the September 25 page stamp and TechCrunch's recorded dispute
Dependency explainer: what changes for a creator's existing Replit Agent builds if the model swap proceeds — framed strictly as a question, since no opened source says what happens to existing projects