By ByteBot
Replit CEO Amjad Masad took the TechCrunch Disrupt stage on Thursday and told the world his company “no longer works with Anthropic.” Just 17 days before that announcement, Replit’s own engineering blog documented Claude Fable 5.1 running in production alongside GPT-6 Astra. Someone’s story doesn’t quite add up.
What the CEO Said
Speaking at TechCrunch Disrupt on October 16, Masad confirmed Replit has moved its primary model infrastructure to OpenAI’s GPT-6 Sol (Power Mode) and GPT-6 Astra (Max Mode), citing three reasons: cost, performance, and partnership quality. “I think ChatGPT is better now,” he said, pointing to Replit’s internal VBench benchmark, which measures performance on “vibe-coded applications from scratch.” Sol and Astra consistently topped those results, he explained.
The OpenAI partnership itself was announced in August 2026, initially framing it as an additive move — never as a replacement. There was no public indication at the time that Anthropic was being phased out. Masad also emphasized that Replit “does not intend to bet the company on any single model provider,” positioning the shift as multi-model strategy rather than a hard switch.
What Replit’s Engineering Blog Actually Shows
That’s where it gets complicated. On September 29 — 17 days before the TechCrunch announcement — Replit published an engineering post titled “Free the models” documenting its actual production model usage. The numbers: Claude Fable 5 in August, Fable 5.1 and GPT-6 Astra running side-by-side in September, with subagent dispatch rates of 32%, 21%, and 36% respectively. A Replit product employee was also publicly promoting Claude Opus 5.5 around the same period.
Anthropic insiders pushed back on Masad’s framing. According to TechCrunch, people familiar with the partnership said the collaboration was still active and contested the “no longer working with” characterization. TechCrunch requested clarification from Replit; no response had been received at publication time. Replit’s documentation, as of this writing, still lists multiple Claude versions alongside OpenAI options.
So is this a clean break or a strategic pivot framed as one? The engineering data says multi-model was already operational before the announcement. The CEO’s language on stage — “no longer working with” — implies something more definitive.
The Structural Problem Nobody Is Talking About
Beneath the partnership drama sits a more uncomfortable truth: Replit depends on frontier AI labs for the models that power its product. Those same labs now sell products that compete head-to-head with Replit.
Anthropic ships Claude Code. OpenAI ships Codex. Both are agentic coding tools targeting the same developers Replit courts. Switching from Anthropic to OpenAI as a primary partner doesn’t resolve the conflict — it trades one competitor-supplier for another. The structural problem is identical in either configuration.
This isn’t theoretical risk. On June 12, 2026, a US Commerce Department export-control directive forced Anthropic to disable Claude Fable 5 and Mythos 5 globally, overnight, for every customer worldwide. Any developer platform running solely on those models lost functionality with no warning. That event alone validates why multi-model isn’t aspirational — it’s defensive infrastructure.
What This Means If You’re Building on an AI Platform
The Replit story is a useful data point for any developer building a product that depends on a third-party model provider. A few things worth taking from it:
- Multi-model routing is now table stakes. Not because it improves performance (though it can), but because a single export ban, pricing change, or competitive pivot can eliminate your product’s core functionality overnight.
- When the CEO contradicts the engineering blog, read the engineering blog. The September 29 post showed the actual operational reality — multi-model was already live. The announcement on stage was positioning, not a factual end-state.
- Your AI supplier may become your direct competitor. Replit’s predicament isn’t unique. Any platform deeply integrated with a frontier lab faces the same structural pressure. Plan accordingly.
Masad’s multi-model framing is the right instinct. The messy execution — a public announcement that contradicts the engineering reality by 17 days, with insiders disputing the claim the same afternoon — is a reminder that these relationships are complicated, high-stakes, and increasingly competitive. Worth watching regardless of which model you’re building on.
Source: TechCrunch | Additional reporting: Unite.AI, CryptoBriefing
