Cognition

Introducing SWE-2 Flash, our new fast & efficient coding model.

On leading evals, it outperforms all American open-weight models, even ones 3x larger.

To balance transparency and cyber risks, we are building a program to securely share the weights with partners & researchers.

Introducing Devin Review 2.0: the best way to review PRs with Devin.

Now up to 5× cheaper, 1.5× faster, and with new features and integrations.

How should you secure your stack?

Our Head of Security Products @jackhcable shares a practical guide for startups and tech teams of every size.

We've evaluated Gemini 4 Carbon and Gemini 4 Argon on FrontierCode 2.0, our coding benchmark that measure's a model's ability to go beyond the spec.

Gemini 4 Carbon scores near SOTA with a score of 40% and a cost of $4.04 per task, behind Claude Opus 5.5, GPT-6 Astra, and Claude Sonnet 5.5.

Gemini 4 Argon scores 26.3% with a cost of $6.41 per task, ahead of Muse Wonder 1.0, but behind Claude Haiku 5.5 and Grok 4.7.

See the leaderboard with full scores: cognition.com/frontiercode

Today, Cognition is acquiring @corridor.

Securing the world’s software is the most important problem in technology today.

Let’s get to work!