Replit CEO says it stopped working with Anthropic; the same story carries a denial

Replit sells a browser-based AI coding platform: you describe what you want, and its Agent builds and deploys the app. On October 15, 2026, CEO and co-founder Amjad Masad said on the Disrupt Stage at TechCrunch Disrupt in San Francisco that the company is no longer working with Anthropic. The official agenda puts his session, “Everyone Can Build Software Now. What Happens Next?”, at 3:15–3:45 PM PT that day. Only TechCrunch has his words on the record, in a story published October 16 — and that same story carries an unnamed person familiar with the arrangement saying the partnership “was still in place.” TechCrunch writes that it reached back out to Replit for clarification and would update if it heard back; no Replit response appears in the piece.

Key points

• Masad gave three reasons — cost, partnership and performance. He said “I think ChatGPT is better now,” and that Sol and Astra consistently perform better at building vibe-coded applications from scratch, based on an internal benchmark plus a public one the company publishes.
• The rebuttal sits in the same TechCrunch report: one unnamed person said the partnership was still in place, and the article cites a late-September blog post listing Claude Fable 5 and Fable 5.1 in Replit Agent’s production environment, plus a Replit product employee posting about Opus 5.5 around the same time.
• Masad’s growth numbers from the same stage: enterprise was about 2% of revenue last year; this year enterprise contracts are 15% and B2B is more than half. He declined to share ARR — “not sharing ARR numbers just yet.”

What was said

TechCrunch senior writer Lucas Ropek recorded Masad saying the decision was made “for a number of different reasons, including cost.” On model quality: “I think ChatGPT is better now,” 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.”

That quote has no second source. TechCrunch is the only outlet that was in the room and wrote it up. On October 17 I tried to pull the automatic captions from TechCrunch’s own recording of the session (grkRhVKcxiE) to check the wording line by line; the caption endpoint returned nothing, so the quote stays attributed to TechCrunch and is not presented here as verified from video. The other outlet that prints the quote is Unite.AI, in an October 16 piece that renders it as “I think GPT is better now” — not “ChatGPT.” That site states in its own author bio that Aiden Cross’s articles are AI-generated and reviewed by an editor. Where the two quotes disagree, this entry follows TechCrunch and keeps the discrepancy visible.

TechCrunch does not name the public benchmark. Replit’s own most recent published numbers are from its September 29 engineering post, “Free the models,” covering DeepSWE v1.1 and Terminal-Bench 4.0. The post does not connect those results to anything Masad said at Disrupt; the only link between them is the order in time.

The counter-evidence

The counter-argument turns on dates. The blog post TechCrunch cites is the September 29 one, and it says production ran “one week per model: Fable 5 in August 2026, Fable 5.1 and Astra in September.” Replit’s docs page for the model selector also still lists Claude; when I fetched it on October 17, the page’s own dateModified was 2026-09-25T09:31:53Z. Both traces sit before October 15. They establish that Claude ran in August and September, not that it was still running on the day Masad spoke — the same line TechCrunch draws with “as recently as a few weeks ago.”

Masad himself framed the matter as supplier risk management rather than a breakup. Crypto Briefing’s headline says the company is “no longer working with Anthropic,” but its body states plainly that Replit has not publicly announced any termination, that the Claude integration “remains active” and has powered Agent since 2024, and it reads the session as “less about ending relationships and more about making sure no single relationship can end Replit.”

There is a second, tighter tie between the two companies. A Replit blog post on June 17 announced Replit running inside Claude: design an app in Claude Design in natural language, then hand the task to Replit through the official Replit Connector to build and ship it. That product line never came up at Disrupt, but it shows the interface between the two companies is not only model calls.

The facts and numbers

The model list below is Replit’s public documentation as of October 17 (dateModified 2026-09-25). Seven Claude models appear across the two manual modes:

Mode	Models
Power	GPT-6 Sol, GPT-6 Luna Fast, Claude Sonnet 5, Claude Sonnet 4.6
Max	GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5, Claude Opus 4.8, Claude Opus 5, Claude Opus 5 Fast, Kimi K3

The Sol and Astra Masad named map to GPT-6 Sol (Power) and GPT-6 Astra (Max). Replit’s September 29 post also published delegation behavior from live traffic, at medium reasoning effort:

Model	Production window	Turns dispatching a subagent	Turns handing work to a general worker	Dispatches returning to an existing subagent
Claude Fable 5	August 2026	32%	0.9%	17%
Claude Fable 5.1	September 2026	21%	2.3%	29%
GPT-6 Astra	September 2026	36%	20%	42%

The benchmark table from the same post (means of four repetitions, Replit Agent in Max mode with Astra as the core loop):

Benchmark	Replit Agent	Astra alone, low effort	Astra alone, xhigh	Sidekick architecture
DeepSWE v1.1, 113 tasks	72% at $2.11/task	67% at $1.60	74% at $4.43	61% at $1.34
Terminal-Bench 4.0, 63 tasks (3 GPU tasks excluded)	49% at $2.53/task	42% at $2.25	60% at $5.86	33% at $1.84

Every one of those figures is a vendor claim. The Astra baselines come from published mini-swe-agent results on each leaderboard; the sidekick architecture is Replit’s own control. Nothing here has been independently reproduced. The post’s claim that the code quality of Astra and Fable 5.1 “let us use our code-review subagent less, with no drop in our eval scores” is self-reported too.

On revenue, Masad said enterprise contracts are 15% of revenue, B2B more than half, and last year’s enterprise share about 2%, and he withheld ARR. The company’s own March 11 funding post is bigger: $400 million raised at a $9 billion valuation, a 3x jump in six months, with the claim of being “on track to hit $1 billion in run-rate revenue by the end of 2026.” Crypto Briefing attributes the $3 billion-to-$9 billion climb to a “$250 million funding round,” which does not match Replit’s own numbers — the $250 million round was September 2025 at $3 billion; the March 2026 round was $400 million.

What others say

Crypto Briefing reads the story structurally: model providers own the most expensive and hardest-to-replicate layer, application companies own the user relationship and the workflow, which is why vendor lock-in has become a board-level concern across AI software. Masad’s emphasis on flexibility, in that reading, is a hedge against exactly that, and it may resonate with enterprise buyers who worry about the same concentration. For investors, the piece flags the flip side: a company whose flagship product leans on a third party’s models carries exposure a traditional software firm would not, and a $9 billion valuation prices in a lot of continued execution.

Replit’s multi-model line has a product-side record. An August 26 post, “Intelligent Model Routing,” says the routing delivered the same output quality “at 65% lower cost than the previous version of Max Mode,” and that enterprise administrators can define an approved model set per workspace for the system to choose from. That is the direct context for Masad’s “cost” reason — and it is a post that does not mention Anthropic at all.

Unite.AI adds material TechCrunch does not have: Masad answering “Yeah. I mean, yeah” when asked directly, Replit having been one of Anthropic’s top five customers, net revenue retention above 400%, an IPO in the company’s future, and turned-down acquisition approaches. None of that is verifiable in TechCrunch’s text or in Replit’s own materials, so this entry records it only as claims from Unite.AI’s AI-generated piece. Its account of the August 18 OpenAI partnership does check out: Replit’s own announcement says Free Mode is powered by OpenAI’s GPT-5.6 Luna, that Core subscribers can create up to 30x more for their $20 monthly subscription with up to 30 hours of chat, and that usage limits reset every five hours.

Neither party has a public position. Reading anthropic.com/news on October 17 turned up no statement about Replit, and Replit has issued none beyond the onstage remarks and the unanswered request for clarification. Anthropic’s share of small-business spending is still climbing — we covered UK SME AI spending data showing Anthropic taking 54.1% with Replit at 3.9% — and that neither proves the partnership is live nor that it has ended.

Our take

For anyone building on Replit or a platform like it, the useful part of this story is not who won. It is that an application company’s dependence on model suppliers is now a risk people discuss in public. Masad split his reasoning into cost, partnership and performance, and only cost is externally observable: unit prices, cost per task, routing policy — all visible in a bill or a docs page. Performance and partnership are internal judgments, and Replit’s published benchmarks are self-reported, so they cannot validate the claim that Sol and Astra are better.

Three things remain genuinely unknown. Whether the partnership has ended cannot be verified, because neither party will speak and the only counter-evidence comes from one anonymous person. Which models ran in Replit’s production after October 15 is not documented anywhere public; the most recent first-hand record stops in September. And if the relationship really is over, whether Replit pulls those seven Claude entries from its docs is the easiest signal to watch, because it would be a visible deletion.

If your work depends on a specific model’s behavior — how a long task delegates to subagents, say — do not rely on a vendor’s partnership announcements. Open Replit’s model-selector docs page, check which models are listed for the mode you use, and check again later, because that page’s dateModified tracks reality more closely than any interview. The Opus 5.5 a Replit employee was promoting on X is the same model that took first place on Arena’s agent leaderboard in early October; the docs list and the leaderboards move together.