Replit makes ChatGPT the default model over Claude
Replit, the platform for building apps with AI, is no longer working with Anthropic, and its coding agent now uses ChatGPT from OpenAI by default. Founder and CEO Amjad Masad announced this on October 16 at TechCrunch Disrupt and clarified afterwards that Claude models remain available — but the default has flipped. And most users never change the default.
What Masad actually said
He named three reasons: cost, the partnership terms, and quality. "I think ChatGPT is better now," Masad said. According to him, both Replit's internal benchmark and a public one consistently show that OpenAI's Sol and Astra models are better at the platform's core job — building applications from scratch. Not long ago Replit was a top-five customer of Anthropic, and that partnership reportedly helped the company grow its revenue roughly tenfold.
A telling detail: "no longer working" is louder than reality. Sources familiar with the deal told TechCrunch the partnership is still formally in place, and as recently as late September Replit promoted Claude Fable 5 and 5.1 models as powering its agent in production. The change is not a hard break — it is the default. OpenAI is now the preferred partner for Replit Agent.
Why the default is the whole story
Most users stick with what ships as the default. Flipping it can quietly shrink Replit's volume at Anthropic — and materially move Anthropic's numbers, for which Replit was a large customer. For the coding-agent market it is another signal: model choice is money, and platforms have started switching users to whatever is cheaper and faster without asking permission.
Replit's own trajectory matters too: enterprise was about 2% of revenue a year ago, roughly 15% now, with more than half of revenue from B2B — and Masad expects over half to come from enterprise and government customers in 2027. With that profile, "whose engine is under the hood" stops being a technical footnote.
Numbers from our side
Replit did not disclose savings, and the internal benchmark is not public — honest to note. But the "no single default model" picture shows up in our data too. By anonymized aggregate counters on NeuralSpace over 30 days (Sep 18 – Oct 18, 2026), 296 chat sessions spread out like this: GPT-6 Sol — 50 sessions, DeepSeek V4.1 Flash — 44, Claude Sonnet 5 — 38, GPT-5.4 mini — 35, with a long tail after that. People mix models per task instead of settling for the first one.
Model quality is measurable in practice, not just in words. Per platform measurements over 90 days (Jul 20 – Oct 18, 2026), the share of successfully completed generations was: GPT Image 2 — 93.2% (3,202 generations), Seedream 4.5 — 90.8% (1,412), Kling 3.0 — 91.6% (297). Those are the kind of aggregates — with a period and sample size — worth asking of any service that "just changed its default".
What it means for developers
If you build on someone else's model, the default under you can change overnight — design for multi-provider and manual model choice. If you are a user, take the choice into your own hands: our chat with models and code mode let you switch models per task rather than per someone's business decision. Source: TechCrunch.
FAQ
Does Claude still work in Replit? Yes, it can still be selected manually. Only the Replit Agent default changed — it is now ChatGPT.
Why did Replit drop Anthropic? Per the CEO: cost, partnership terms, and benchmark performance. No public savings figures were disclosed.
What does it mean for Anthropic? Replit was a top-five customer. A default flip, given user inertia, means materially lower volume even if the formal partnership survives.