Every AI coding tool you use right now — Cursor, Copilot, Windsurf — runs on someone else's brain. You type code, the editor sends it to Claude or GPT through an API (a pipe that connects your editor to a remote AI), and a model built to write poetry, plan vacations, and also code sends back a suggestion. It works. But it is slow, generic, and you are renting intelligence from companies that might eat your lunch tomorrow.

Windsurf — the AI-native IDE that started life as Codeium, then got acquired by Cognition in 2025 — decided renting was for suckers.

On October 29, 2025, Cognition released SWE-1.5, the latest version of their custom model family built specifically for software engineering. Not a fine-tune (tweaking an existing model with extra training data). Not a wrapper. A frontier-size model with hundreds of billions of parameters (the internal knobs an AI uses to make decisions). Cognition trained it end-to-end with reinforcement learning (an AI teaching method where the model learns by trial and error) on real coding tasks.

On SWE-Bench Pro — the industry-standard benchmark of 731 challenging coding tasks across 41 code repositories — SWE-1.5 scored 40.08%. Claude Sonnet 4.5 scored 43.60%. Close enough to raise eyebrows, especially given what comes next.

Windsurf partnered with Cerebras, a chip company that builds specialized AI hardware, to serve SWE-1.5 at up to 950 tokens per second. A token is roughly three-quarters of an English word — so 950 tokens/sec means the model generates text roughly 13x faster than Sonnet 4.5 and 6x faster than Haiku 4.5. Cognition also rewrote their internal lint checking (automated code error detection) and command execution pipelines, cutting up to 2 seconds of overhead per step in agent sessions — those iterative loops where AI reads code, runs tests, reads errors, and fixes things.

The practical result: agent workflows that take minutes on Sonnet finish in seconds on SWE-1.5. For developers running fix-test-fix-test cycles all day, that speed difference compounds into hours saved per week. The feedback loop between "fix this" and "it is fixed" shrinks from coffee-break to blink-and-done. For a cat who values naps and efficiency, deeply appealing. 😸

On December 24, 2025, Wave 13 shipped SWE-1.5 as the default model — free for all users for three months. That free period ended on March 27, 2026, when Windsurf rolled out a new quota billing system. Classic dealer strategy: first taste is free, then you are hooked on the speed and $30/month suddenly feels reasonable. Smart. Devious. 😹

But here is the part nobody wants to talk about. Cognition trained SWE-1.5 specifically on agentic coding tasks — then tested it on a benchmark that measures agentic coding tasks. Scoring well on the exam you studied for is expected, not impressive. The real question is whether SWE-1.5 holds up on messy, real-world codebases that look nothing like training data.

And there is a bigger concern. If every IDE company trains its own proprietary model, we end up with walled gardens. Your IDE choice determines your AI model. Your model determines your IDE. Vendor lock-in with extra steps. At least when everyone used GPT-4, you could switch editors without retraining your instincts around a completely different model's quirks. 😾

Still — training your own frontier model instead of paying Sam Altman rent? That is the correct strategic move. Every IDE company sending API calls to OpenAI funded its future competitor. Cognition said "no thanks" and built their own brain. The benchmarks back it up. The speed backs it up even more.

Two years ago, every AI coding tool was a thin wrapper around OpenAI's API. Today, Windsurf trains its own models, Cursor runs custom fine-tunes, Copilot has moved beyond raw GPT completions. The IDE layer is becoming the model layer. Competition is good. The developer wins.

The cat watches. 🐈