Replit CEO Amjad Masad said onstage at TechCrunch Disrupt 2026 that Replit no longer works with Anthropic, linking the move to cost, partnership and performance. The session recording was published by TechCrunch on October 16, 2026. Masad added that Replit was one of Anthropic's top five customers, which gives the split weight for both sides. For business users, the message is direct: a leading coding platform has picked its model supplier on commercial terms.
Why Replit switched to OpenAI models
Masad confirmed the break when asked whether the OpenAI partnership announced in August 2026 meant an end to work with Anthropic. His answer was short: "Yeah. I mean, yeah." He then said OpenAI models now perform better for Replit workloads, stating that GPT is better now. As evidence he cited Replit's internal VBench benchmark and a public benchmark published by the company. OpenAI's Sol and Astra models, he said, consistently performed better at building vibe-coded applications from scratch.
The partnership dates to Replit's August 18, 2026 announcement of Free Mode, powered by OpenAI's GPT-5.6 Luna. Replit said Core subscribers could create up to 30 times more than before for a $20 monthly subscription, including up to 30 hours per month of chat. Everyday tasks in Free Mode stopped consuming credits. The same announcement defined Power Mode, formerly Economy Mode, as using cost-optimized models, while Max Mode uses higher-performance models with usage limits resetting every five hours.
Current model-selector documentation still lists a broad mix, including Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5, Claude Opus 5, Claude Opus 5 Fast, Claude Opus 4.8 and Claude Sonnet 4.6 as Max Mode options. Alongside them are GPT-6 Astra, GPT-6 Luna Fast and Kimi K3, with GPT-6 Sol offered in Power Mode. Free Mode always uses intelligent model routing with no manual selection, while Fast variants require a Pro or Enterprise plan. A September 29, 2026 engineering post documented Claude Fable 5 in Replit Agent production in August and Fable 5.1 alongside GPT-6 Astra in September.
What the split means for AI platform buyers
For companies building software with AI agents, the production data in that engineering post matters more than the headline. It shows Fable 5 dispatching a subagent on 32% of turns, Fable 5.1 on 21% and Astra on 36%, each at medium reasoning effort. Replit Agent in Max mode with Astra as core loop scored 72% at $2.11 per task on DeepSWE v1.1 and 49% at $2.53 per task on Terminal-Bench 4.0. That setup beat a sidekick architecture by 11 and 16 points. Small teams get a clear price-to-quality anchor per task, while large buyers can compare orchestration designs.
Enterprise controls and economics form the second practical layer. Masad said enterprise contracts account for about 15% of Replit revenue in 2026 against about 2% in 2025, with more than 50% of revenue now B2B. He put net revenue retention above 400%, saying adopters spend 400% more within 12 months, and said Replit has been gross-margin positive. He also listed pre-deployment scans, post-deployment penetration-test scans, supply-chain monitoring and a one-click kill switch, plus one-click restore from snapshots after a 2025 incident in which an agent deleted a database.
The marker to watch is whether OpenAI remains the better partner for now, in Masad's wording, through measurable renewals. Masad declined to share ARR figures, said there are no immediate plans to raise funding, pointed to a future IPO and said acquisition approaches were turned down. If enterprise share keeps rising from 15% and B2B stays above half of revenue while Free Mode and Power Mode pricing hold, the supplier switch will look durable. If model rankings on VBench shift back, multi-model routing may return to the center.
