The two-vendor coding-agent split is now real — quota-bundled autonomous engineering vs per-token model routing
Devin 3 hits 90% SWE-bench Verified. Cognition completes Windsurf at $250M. Cursor Composer 2.5 ships Build in Parallel. The agent-IDE market just settled into a clean two-vendor split with materially different pricing models. Both are defensible. Procurement teams can finally pick on operating model, not capability.
The market structure clarifies
Devin 3 clears 90% on SWE-bench Verified — the first autonomous engineering agent to hold that threshold consistently. Cognition closes its $250M acquisition of Windsurf (the remaining stake after Google's $2.4B acqui-hire of the founders), and Windsurf 2.0 ships with Devin Cloud and Devin Terminal CLI bundled inside the IDE at $200/month Max tier.
On the other side, Cursor Composer 2.5 (from the AM cycle) holds the per-token-pricing lane at $0.50/M input / $2.50/M output, with MS Teams integration and Build in Parallel.
The two-vendor split
The market has settled into two operating models:
- Quota-bundled autonomous engineering (Windsurf+Devin): Fixed-cost subscription, autonomous-engineer-equivalent budgeting, predictable spend. Enterprise IDE buyers who would otherwise quibble over per-task billing absorb $200/month as a developer-tooling line item.
- Per-token model routing (Cursor): Pay-per-use elasticity, model-routing flexibility, deepest IDE muscle-memory. Teams that want to dial routing by workload and accept variable monthly cost go this direction.
Why both are defensible
The two models serve different procurement preferences, not different capability tiers. A team that wants subscription predictability gets it from Windsurf+Devin without losing capability — Devin 3 at 90% SWE-bench is competitive with Cursor's Opus 4.7 / GPT-5.5 routing on per-task benchmarks. A team that wants per-token elasticity gets it from Cursor without losing autonomous-engineering capability — Build in Parallel covers the autonomous lane.
The agent-IDE market has matured past the 'find the best model' phase. The choice now is 'which operating model fits your procurement preferences.' That's a healthier maturity than the field had a year ago.
The MCP-protocol substrate
MCP server registry continues exploding — over 800 production MCP servers indexed. Both Cursor and Windsurf+Devin consume MCP servers; tool integration is increasingly portable across the two-vendor split. The market structure assumes MCP as the tool-access substrate; vendor lock-in lives at the model-routing and IDE-UX layers, not at the tool-integration layer.
The forward read
Both vendors will keep scaling ARR through 2026 H2 (Cursor at $1.2B ARR, Claude at $2.5B annualized). The interesting question is whether a third operating model emerges — possibly a hybrid pay-as-you-grow tier that bridges the two — or whether the market stays cleanly bifurcated. The 2027 watch.
Lushbinary — AI coding agents comparison → · Kingy AI — practical map → · Shareuhack — Cursor vs Claude Code vs Windsurf →