Replit Just Quit Anthropic — The First Dev-Stack Defection Tests a $965B IPO Thesis

Replit co-founder and CEO Amjad Masad told an audience at TechCrunch Disrupt on Oct. 16, 2026 that the company is no longer working with Anthropic. The platform, which he said has 50M users and is on track to hit $1B in run-rate revenue by year-end, previously ran on Anthropic's Claude models through Google Cloud's Vertex AI and most recently shipped Claude Sonnet 5.5 and Claude Fable 5/5.1 in its Agent production environment as recently as late September.

The \"better partner\" is OpenAI. Replit and OpenAI publicly linked up on Aug. 19, 2026 when Replit launched \"Free Mode\" powered by OpenAI's GPT-5.6 Luna, and the platform's Oct. 2, 2026 changelog listed GPT-6.1 Sol as its Max Mode default. Masad added an internal benchmark plus a public benchmark to justify the switch: in his framing, OpenAI's Sol and Astra \"consistently perform better on building vibe-coded applications from scratch.\"

Replit is the second dev-stack to publicly walk away from a frontier lab this quarter. On Aug. 14, 2026 SpaceX closed a $60B all-stock acquisition of Anysphere, parent of Cursor; two weeks later OpenAI said it would terminate its model contract with Cursor on Nov. 12, 2026, citing concerns about \"contract/security of proprietary technology.\" OpenAI's models were about 5% of Cursor's user traffic. Anthropic, which had been Cursor's primary model supplier, publicly responded: co-founder and CCO Tom Brown said the lab will \"continue to increase compute to support Claude models in Cursor,\" and explicitly welcomed SpaceX ownership.

Read the table row by row and the pattern is clear: when a frontier lab crosses into a gatekeeper's product lane, the gatekeeper reroutes. Cursor chose Anthropic because OpenAI's Claude Code competing product muddied the relationship. Replit chose OpenAI because Anthropic's Claude Code and Claude Design ship into the same vibe-coding slot.

Anthropic closed its Series H on May 28, 2026 at a $965B post-money valuation, with $65B raised led by Altimeter, Dragoneer, Greenoaks, and Sequoia. By July2026 the company was running at a $65B annualized revenue rate, up roughly 600% from year-end 2025. Q2 2026 revenue alone crossed $11.5B versus $787M a year earlier — Anthropic's first quarter of positive adjusted operating income. Roughly 80–85% of that revenue comes from enterprise and developer customers paying for API access directly or via Amazon's AWS Bedrock or Alphabet's Vertex AI.

The non-obvious hit isn't the immediate revenue — Replit was a small slice of Anthropic's $65B run-rate. It's the precedent. Replit is the first dev-stack to publicly cite \"partnership\" alongside cost and performance, and to frame Anthropic as a competitor rather than a partner. That's the language an S-1 can't easily explain away, and it's the language that gets echoed by Vercel and Lovable when they negotiate.

Strip the model decision away and Replit's strategy is clearer. The company raised $400M at a $9B valuation on Mar. 11, 2026 (3x in six months), hit an estimated $525M ARR in April 2026, and told investors it was on track to $1B run-rate by year-end. Enterprise revenue jumped from roughly 2% of the mix last year to about 15% this year, with B2B now over 50% of revenue — UKG alone reported a 400% increase in product-feedback throughput on the platform.

The enterprise gateway is Microsoft Fabric. Replit announced a collaboration with Microsoft in July 2025 and deepened it through Microsoft Build 2026, with deploy-to-Fabric tooling letting Replit-built apps land inside an enterprise-managed Fabric tenant. The Self-Driving Company blog post Masad published in July 2026 also described a \"harness\" architecture that routes between models — GPT-6 Astra delegating to cheaper subagents, etc. — which means the model decision is reversible at the routing layer. That's a structural moat: a customer-facing reason to stay on Replit regardless of which model is winning that week.

Model-switching by distribution gatekeepers routes compute and revenue up the stack. Four listed names sit at the points of leverage:

Short-term (days to quarters), the catalyst path is concrete: Anthropic's next earnings or pre-IPO investor update will disclose whether Replit's exit is one-off or symptomatic. The Cursor/Anthropic deepening is a useful read — Anthropic publicly committed to add Claude compute to Cursor the same week Replit quietly stopped using it. That bifurcation (more compute to Cursor, less to Replit) is itself the signal.

Long-term (one to three years), three things to watch: (1) whether Replit's enterprise revenue actually hits the 15%-of-mix target on Microsoft Fabric, which would validate the model-agnostic positioning as a real moat; (2) whether Anthropic's IPO timeline slips past 2026, since a $965B private valuation is only supportable if enterprise API lock-in is preserved; (3) whether Vercel, Lovable, and any unannounced Replit-tier platform follow Replit's lead. Two defections in 60 days is anecdote; three is a market structure.

For investors, the cleanest read is that model leadership is necessary but no longer sufficient — distribution gatekeepers (Replit, Cursor, Vercel) have enough volume, brand, and enterprise footprint to negotiate. That shifts pricing on Claude tokens, it shifts the IPO narrative for Anthropic, and it concentrates enterprise compute spend with the hyperscaler that hosts the winning model. As of October 2026 that is Microsoft (GPT-6.1 Sol default on Replit, OpenAI backer, Fabric gateway) and secondarily Alphabet (Replit's original Vertex AI partner). Amazon is the cleanest pure short on a continued gatekeeper drift away from Anthropic; NVIDIA is the structural long on every branch of the decision tree.