When Cognition closed its acquisition of Windsurf in July 2025, the move shifted the center of gravity in AI coding from autocomplete to autonomy. Devin gained an editor surface; Windsurf gained an agent that finishes pull requests. By September, Cognition was repriced to $10.2B; by April 2026, the company was reportedly in talks at $25B. Cursor answered with its own escalation — talks at $50B pre-money in April 2026, against a $2B ARR run-rate. The 40 startups in this slice account for $14.5B in cumulative funding, and Cursor's $7.57B Series D alone is more than half of it.
Where the benchmarks landed
Claude Opus 4.5 and Opus 4.7 sit on top of the Aider Polyglot leaderboard at 89.4% — the canonical multi-language coding eval that runs 225 Exercism problems across C++, Go, Java, JavaScript, Python, and Rust. SWE-bench Verified shows similar standings: Claude Opus 4.7 (Adaptive) at 87.6%, GPT-5.3 Codex at 85%. On the open-weight side, GLM-5 hit 77.8% on SWE-bench Verified — close enough that BYO-model agents like Cline and OpenCode are now meaningfully competitive.
Who is shipping and at what price
Cursor, Cognition (Devin), Replit (Agent, $9B Series F), Poolside ($500M Series B at $3B), Augment Code ($252M Series B), and Sourcegraph (Cody, $228M Series D) are the funded independents. Tabnine ($55M Series B at $1.5B) holds the enterprise-private niche. The free-or-bundled tier is Microsoft GitHub Copilot, Google Gemini Code Assist, Amazon Q Developer, Amazon Kiro, and JetBrains AI Assistant — distribution-first and roughly zero-cost if you already pay the parent cloud. Cline (open-source, $4M seed at $110M) and OpenCode are the BYO-model wedge.
What changes the picture from here
Frontier coding models commoditize fast — every six weeks brings a new Aider Polyglot leader, and per-token prices keep falling. The independent thesis depends on the editor surface compounding around the model rather than under it. Cursor's reported $2B ARR is the proof point. Devin's autonomous-agent pricing — billed against task throughput, not seats — is the alternative. The question for the next 12 months is whether GitHub Copilot's free tier with frontier-grade quality compresses everything in the middle.
