On Thursday October 16, 2026, Replit co-founder and CEO Amjad Masad sat on a TechCrunch Disrupt stage in San Francisco and was asked whether the OpenAI partnership Replit announced in August meant the end of its relationship with Anthropic. He answered: "Yes. I mean, yes." By Saturday a person familiar with the deal was telling TechCrunch the Anthropic partnership was still in force, Replit's model picker was still listing Claude variants, and Masad had moved to a very different sentence: "Replit remains model-agnostic." TechCrunch kept the URL — replit-is-no-longer-working-with-anthropic-ceo-says — but softened the on-page headline. Everyone involved spent the weekend rebuilding a sentence that had been said on stage in front of cameras.

Masad is not wrong about what Replit decided. The walk-back is defending the wrong number.

What he said while the microphone was on

Masad's framing on stage was unusually direct for a CEO talking about a hyperscaler-tier vendor. On cost, partnership terms, and performance, he said he thought "ChatGPT is better now" and that "consistently, Sol and Astra perform better on building vibe-coded applications from scratch." He cited an internal benchmark and one public benchmark. He named OpenAI as "our preferred partner for powering Replit Agent." Asked whether the shift had contributed to Anthropic's slowing growth, he confirmed Replit had been "among the top five" customers Anthropic reports by revenue.

The business numbers he volunteered the same day sit under the model switch. Replit's enterprise revenue climbed from roughly 2% of total revenue in 2025 to about 15% in 2026, with more than half of all revenue now B2B. Net revenue retention on business customers runs "over 400%," a number TechCrunch could not independently verify. ARR, as always, was declined. By 2027 Masad expects more than half of revenue to come from enterprise and public-sector contracts. In that mix, the model powering the default agent isn't a line item — it's the product.

Masad also walked through Replit's own usage data from an engineering post dated September 29: subagents were dispatched in 32% of Claude Fable 5 interactions, 21% of Fable 5.1, and 36% of GPT-6 Astra. On Replit's own numbers, Astra was pulling ahead on the one behaviour the agent product monetises hardest.

The walk-back, in order

Inside 48 hours three things happened in parallel.

First, Masad told reporters after the panel that "Replit remains model agnostic," and that users can still pick Claude from the model selector. This is true: the selector still carries Claude Fable 5.1 and Opus 5.5, and the engineering post from September names both as in-production options.

Second, a person familiar with the deal — identified by TechCrunch only as "familiar with the deal" and described in some aggregator write-ups as plural "insiders" — pushed back that the Anthropic partnership "is still in place," pointing at the September production footprint as evidence the relationship isn't over.

Third, TechCrunch edited the on-page headline from "Replit is no longer working with Anthropic, CEO says" to a softer form, but left the URL slug intact. Someone made a decision to walk the headline without erasing the record.

That sequence tells the actual story. The partnership isn't dead; the default is. Masad's on-stage answer was about who powers Replit Agent by default, which is also the question whose answer is worth real money. The post-event clarification is about what shows up in the picker, which is the question whose answer costs almost nothing. Both are true. Only one of them moves the invoice.

Defaults are the only number that matters

Replit's agent product is sold to users who do not pick models. The 2% → 15% enterprise revenue jump is not a story about power users switching to Opus 5.5 — it is a story about seats, administrators, and policy templates. Enterprise seats default-on whatever the vendor's default is. If Replit's default moves from Claude to Sol/Astra, Anthropic's share of a top-5 account moves with it, picker or no picker.

How much? Nobody at Replit will say, and nobody at Anthropic will either. But a top-5 customer losing default allocation to a competitor is the kind of event an underwriter notices. TechCrunch's October 20 story "Anthropic's cloud partnerships are getting messy," based on a leaked S-1, noted that two customers account for roughly a fifth of Anthropic's revenue. The S-1 does not name them. If Replit was in that stack a quarter ago and isn't by November 9, when the IPO road show is expected to begin, the gap between what the S-1 says and what the current quarter looks like widens on a schedule that lawyers find unwelcome.

This is the third enterprise pullback in two weeks

On October 5, The Information reported that Microsoft had cut per-engineer Claude budgets inside its cloud division from $100,000 to $10,000 a month, and that Meta had halved its Claude Code user base. The Loop covered that story on October 6, framed as partner-becoming-competitor on a schedule. Nine days later Microsoft Security EVP Hayete Gallot published "Securing the agentic insider," with the line "frontier models, regardless of their origin, cannot be trusted." That ran October 20 on The Loop.

Now Replit, October 16, in the middle of the sequence. Three top-tier customers in two weeks visibly reorganising their Claude posture — one on budget, one on security framing, one on default allocation. Each has a plausible local reason. Taken together they are a pattern the pre-IPO window cannot afford.

The Loop's October 16 piece on Anthropic's Q3 noted that sequential growth fell from +140% to +45% in a single quarter and that the "annualized run-rate" number conspicuously left the deck. The Replit default switch is one of the lines on the chart that produced that deceleration. Masad happened to say it out loud, from a stage, with lights on.

What Anthropic has said

Nothing. TechCrunch asked. There is no on-record statement, no "we remain committed to Replit" line, no counter-benchmark. In an ordinary vendor-relationship story that silence is unremarkable. In a story where the counterparty is five business days from a public S-1, the silence is a choice. The quietest possible response from Anthropic is also the one that gets re-quoted inside every underwriter pitch between here and November 9.

What this means / what to watch

1. The model-mix disclosure. Replit's monthly engineering posts have been publishing per-model usage data. The next one is the actual answer. If Astra's share goes from 36% to the high 60s while Claude's share collapses, "model-agnostic" is a sentence a lawyer wrote. If Claude's share holds above 30%, Masad's stage line was theatre.
2. The S-1 customer-concentration language. Anthropic's confidential S-1 is operationally due to go public around October 23, five business days before a November 9 road-show start. The "two customers represent roughly 20% of revenue" line in last Monday's TechCrunch piece names nobody. The public prospectus will have to. Replit being named — or conspicuously not — is the datapoint.
3. OpenAI's price moves. Sol and Astra pricing has not moved since launch. If Replit's switch becomes a case study inside OpenAI's enterprise sales motion, expect a Sol/Astra enterprise tier announcement before the end of Q1 2027. The sales story writes itself: "we are the vibe-coding default now, and the top-5 customer that used to be Anthropic's agrees."
4. The next defector. Microsoft, Meta, Replit. The pattern that showed up in two weeks will either stop in two weeks or produce a fourth before the IPO. If a fourth top-twenty customer flips its default between now and November 9, the Anthropic story in the public prospectus is no longer about Q3 growth deceleration — it is about customer-concentration risk running in real time while underwriters are pricing the book.
5. Anthropic's response window closes on Oct 23. Any public pushback from Anthropic — a counter-benchmark, a Replit-specific statement, a competing enterprise-coding customer announcement — has to land before the S-1 opens. After that, the quiet period applies and the company cannot talk. The silence from October 16 to October 20 was tolerable. The silence from October 21 to October 23 is a decision.

Masad said the thing a CEO is not supposed to say about a top-5 vendor in front of cameras. The weekend was then spent writing a different sentence that is also true. Both can hold at the same time only because "we no longer work with them" and "they are still in our model picker" answer different questions. The one Masad answered on stage is the one with a dollar sign in front of it.

The one Replit's communications team answered after is the one with a lawyer behind it.