What AI model does it use?

What AI model does Devin use?

Devin, Cognition's coding agent, doesn't hide its model behind a brand name the way some assistants do. You can choose it yourself, and the docs say which one Cognition would pick for you. That answer changed on September 10, 2026, when Cognition released SWE-2.

The recommended model for Devin Cloud

Devin Cloud is the version that works on its own virtual machine while you do something else. When you start a session from the terminal, the /model command sets the model, and the docs say plainly: "Choose the model. SWE-2 is recommended."

SWE-2 is Cognition's newest in-house model, from the SWE family it has been building for a while (the models page also lists SWE-1.7, SWE-1.7 Lightning, SWE-1.6 and a small SWE-1-mini for tab autocomplete). Cognition's launch post puts SWE-2 at 50.0% on its FrontierCode 1.1 benchmark, "within one point of Fable 5.1 while being 64% cheaper". That's Cognition's own measurement, and I haven't seen it repeated independently.

SWE-2 starts from someone else's model

Further down, the launch post says "SWE-2 is post-trained from Kimi K3, a 2.8T-parameter model that had already undergone extensive RL for agentic coding." Kimi K3 is an open-weights model from Moonshot AI. Cognition took it and did more reinforcement learning on its own coding tasks, which it says added five to six points on many benchmarks.

So "in-house" here means Cognition trained the final model, not the base it grew from. Cursor did the same thing with Composer and Kimi K2.5, which makes Moonshot a quiet supplier to two of the best-known coding tools.

Fusion and Adaptive in Devin Desktop

Devin Desktop is what Windsurf became: Cognition renamed its editor on June 2, 2026. There, and in the Devin CLI, the recommendation is different. "For most users, we recommend Fusion," the Fusion page says. Fusion pairs a frontier "lead" model, which plans and reviews, with a cheaper "sidekick" that writes the code. It's for paid plans only; the docs say "It's not included in free or trial tiers."

There's also Adaptive, Cognition's router, which picks a model per request. Self-serve users pay a flat $0.50 per million input tokens and $2 per million output tokens whichever model it lands on, and it's off by default for enterprise accounts.

Which other models you can pick

Cognition says it supports "the latest and greatest models from Anthropic, OpenAI, Google, and Cognition within minutes of their launch." On September 25 the models data included Claude Opus 5.5, Sonnet 5 and Fable 5.1, OpenAI's GPT-6 and GPT-5.6 lines, Gemini 3 models, Grok, GLM, Kimi K3 and DeepSeek V4. The CLI accepts short names like opus, swe or gemini, which resolve to the newest model in each family, and the docs suggest at minimum "trying swe, gpt, and opus."

SWE-2 is free on self-serve plans through October 15, 2026. After that the models page lists it at $3 per million input tokens and $15 per million output.

What Cognition doesn't say

The docs recommend SWE-2 for cloud sessions, but they don't say which model a session uses if you never touch /model. Cognition also doesn't publish which lead and sidekick Fusion pairs by default, or which models Adaptive chooses between.