🔥 Ship your first app in 2 minutes — free Lovable course in CBV Academy
Start Free Course →
Skip to content
codingbutvibes
AI MODELS

Windsurf SWE-1.5: What Cognition Actually Claimed

SWE-1.5 is the coding model Cognition built for the editor that used to be called Windsurf. It is interesting for a reason most coverage skipped: the headline claim is about speed, not about beating anyone on a leaderboard.

Updated October 2026 · 6 min read

Corrected October 2026. This article used to say SWE-1.5 “outperforms GPT-4o” and carried a table of SWE-bench Verified scores for SWE-1.5, GPT-4o, Claude 3.5 Sonnet and Gemini 2.0 Pro. We could not trace those numbers to any published source, and Cognition's own comparison was against Claude Sonnet 4.5 and Haiku 4.5, not GPT-4o. The table is gone. The page also described Cascade and Windsurf's free tier as current; Windsurf is now Devin Desktop, and Cascade has been replaced. Everything below is either Cognition's own claim, attributed to Cognition, or clearly labelled as our reading.

Key Takeaway

Cognition describes SWE-1.5 as a frontier-size model — hundreds of billions of parameters — trained for software engineering and served through Cerebras at up to 950 tokens per second. By Cognition's own comparison that is roughly 6x faster than Claude Haiku 4.5 and 13x faster than Claude Sonnet 4.5. The pitch is near-frontier coding quality at a speed where an agent stops feeling like something you wait for.

WWindsurf

Windsurf

Rising

AI IDE with agentic workflows that plan before they code

✓ Free tier available

Free plan: 25 flow actions/mo, no CC required

Take the free Vibe Coding Stack course →

Paid from $15/mo

What SWE-1.5 is

Most AI coding tools rent a general-purpose model — Claude, GPT, Gemini — and wrap it in an agent. Cognition, the company behind the Devin coding agent, went the other way with its SWE model family: models trained for the specific loop a coding agent runs, which is read the codebase, plan a change, edit several files, run the tests, and go again when something fails.

According to Cognition's announcement, SWE-1.5 is the latest model in that family and the first at frontier scale. It shipped inside Windsurf, the AI editor Codeium built and Cognition took over. That editor has since been renamed: Cognition's own Devin Desktop FAQ says Windsurf became Devin Desktop on 2 June 2026, with the same editor and the same features.

The speed claim, and why it matters more than a score

Cognition partnered with Cerebras, which builds wafer-scale inference hardware, to serve SWE-1.5. The figure Cognition published is up to 950 tokens per second, and it framed that against Anthropic's models: about 6x Haiku 4.5 and 13x Sonnet 4.5. Those are Cognition's numbers about its own product; we have not measured them, and real-world speed depends on load and on how much of a task is spent waiting on tools rather than generating text.

Even so, the idea is the right one to pay attention to. An agent that takes four minutes to make a change gets used differently from one that takes forty seconds. When the loop is fast, you stay in it: you ask, look, correct, and ask again. When it is slow, you hand off a task and context-switch, and every mistake costs you a full round trip. A coding model that is a little less capable but much faster can be the more productive tool for a lot of everyday work.

What about the benchmarks?

Third-party write-ups of the launch quote SWE-Bench Pro scores for SWE-1.5, but they do not agree on the detail, and we could not confirm a figure against Cognition's or Scale AI's own publications. So we are not quoting one. If a number matters to your decision, take it from Scale AI's SWE-Bench Pro leaderboard directly.

Keep in mind what agent benchmarks measure. A score reflects the model and the scaffold around it — the tools it can call, how it is prompted, how many attempts it gets. A model tuned for one agent can look very different inside another. That is especially true of SWE-1.5, which was built to run inside Cognition's own editor.

Where it fits after the Devin Desktop rename

The agent this article was originally written around, Cascade, is gone: coverage of the rename reports it was replaced by a rewritten local agent called Devin Local. Which models Devin Desktop offers, and which one is the default, can change with any update, so check the model picker in the app rather than trusting a launch-day article — this one included. Our Windsurf review covers what else changed.

Who should care

You should, if you already use Devin Desktop or are choosing an agent-first editor, and the thing that frustrates you most about AI agents is waiting for them. A fast model inside the editor it was trained for is exactly the combination worth trying.

It is wrong for you if you want to pick the model yourself and use it anywhere. SWE-1.5 is Cognition's model for Cognition's tools. If you want to switch between frontier models from different labs in one editor, Cursor is built around that; if you want an agent in your terminal, Claude Code, Codex CLI and Gemini CLI are the usual options. Our Cursor vs Windsurf comparison covers that choice in detail.

It is also not a reason on its own to move an existing workflow. If Cursor, Copilot or a terminal agent is working for you, a faster model in a different editor is a reason to try it on a side project, not to migrate a team.

The Bottom Line:

SWE-1.5's real contribution is the argument that speed is a feature for coding agents, not a nice-to-have. Cognition's claim is near-frontier quality at many times the speed of the models it compared against. That is worth testing on your own code. The “beats GPT-4o” framing this article used to carry was never supported, and you should be wary of any page that still repeats it.

💜

Free Course

Build Your First App with Lovable

Hands-on lessons. Build a real project. Lesson 1 is free — no signup needed.

Start Learning Free →

Related Articles