On June 3, 2025, Windsurf CEO Varun Mohan announced that Anthropic had cut the AI coding tool’s first-party capacity to Claude models, with less than five days’ notice. The statement went out through Mohan’s social posts and a Windsurf blog post, and TechCrunch covered it the same day.
Beyond a two-company dispute, the episode is a concrete sample of application-layer risk in AI: when your product is built entirely on someone else’s model, the supplier can change the rules at any time.
Timeline: four months of rapid deterioration
Zoom out, and the cutoff is not a surprise. In February 2025, Anthropic launched its own coding tool, Claude Code. In April, Windsurf hit $100 million in annual recurring revenue and, per Bloomberg reporting, entered talks to be acquired by OpenAI for about $3 billion. In May, Anthropic shipped the Claude 4 models — giving direct access to Cursor, Devin, and GitHub Copilot, but not to Windsurf.
On June 3, the shoe dropped: nearly all of Windsurf’s first-party capacity for Claude 3.x models was cut, on under five days’ notice. In four months, the application company’s relationship with its largest model supplier went from customer to rival. Over the same stretch, rivals like Cursor and Claude Code scaled fast, and the coding tools’ dependence on model suppliers became more sensitive by the month; Windsurf simply dragged that hidden dependency into the open.
First-party capacity slashed on five days’ notice
According to Windsurf, Anthropic eliminated nearly all of its first-party capacity for Claude 3.x models, including Claude 3.5 and 3.7 Sonnet. “We are disappointed by this decision and short notice,” Mohan wrote, adding that Windsurf had been willing to pay for full capacity. Because the remaining third-party capacity was limited, Windsurf warned that Claude 3.x availability could be unstable in the short term.
First-party access means Windsurf bought model capacity directly from Anthropic; third-party access gets the same models through other inference providers. Windsurf used both, but after the first-party cut, capacity ran tight. The interim workaround — letting developers bring their own API keys — is more expensive and more cumbersome for enterprise customers. From an engineering standpoint, third-party inference capacity is not automatically stable: peak-hour latency, rate limits, and billing structures all differ, and for a tool selling instant completions, switching supply paths is not free. Anthropic noted that Claude 4 remained reachable on Windsurf via API keys and its partner ecosystem. One detail worth recording is the notice period: five days. The number will keep getting cited — not because it is uniquely harsh, but because it precisely demonstrates how little bargaining power the application layer has over the model layer.
The business tension: Anthropic also sells coding tools
Beyond timing, the structural cause matters more. Model providers started building applications themselves: Anthropic’s Claude Code, launched in February 2025, grew quickly, and the line between model layer and application layer keeps blurring. When your supplier is also your competitor, supply terms become leverage — the way Claude 4 access was parceled out in May was the tell.
Windsurf’s own situation amplified the drama: $100 million ARR in April, then reported acquisition talks with OpenAI — Anthropic’s rival. Some enterprise users switched to Cursor to keep Claude 4 access, and developer accounts cited by TechCrunch suggest users had come to see model supply on Windsurf as unreliable.
Lessons for teams building on someone else’s model
Three takeaways for AI application developers.
First, model access is a commercial term, not a technical guarantee. Today’s API quota can shrink because of a supplier’s strategy shift, on days — not months — of notice. “Continued supply” and capacity commitments belong in the contract, with a clear-eyed view that the counterparty retains room to change terms.
Second, multi-model routing is no longer an optimization trick but a survival requirement. Had Windsurf spread workloads across more model suppliers earlier, the hit would have been smaller. The same logic applies to any product with a single-model dependency.
Third, when the model provider is also a competitor, the bargaining structure has changed. The supplier benefits from you as long as you sell more tokens for it; once your product threatens its own application ambitions, that interest disappears. Application companies need the technical ability to switch providers at will, and should price this risk into fundraising and valuations. For teams choosing tools today, the checklist this episode leaves behind is concrete: does your current model have alternative sources, does the contract include capacity guarantees, and how fast could your product recover if the supplier changed terms tomorrow?
Sources
- TechCrunch: Windsurf says Anthropic is limiting its direct access to Claude AI models
- Windsurf blog: Anthropic models
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.