← Back to live feed1 story
Spoken commands are now integrated into the Devin AI software engineer via a voice interface that leverages GPT-Live and a new underlying model called SWE-2. This model performs on par with current frontier systems but at a cost up to 70% lower, utilizing reinforcement learning scaled to trillions of parameters. The update is currently available in the Devin Desktop and CLI applications.
Post-trained on Kimi-K3, SWE-2 achieved a 50.0% score on the FrontierCode benchmark, beating GPT 5.6 Sol and Grok 4.6. The model includes effort levels where the medium setting takes 58% fewer turns and costs 81% less than the prior SWE-1.7 iteration. Cognition is offering the model free to Pro, Max, and Teams subscribers for one month.
Sign in to suggest edits
Key sources
- SOURCE@cognition“Your favorite AI software engineer just got a landline”x.com
- SUPPORT@cognition“Try it today in Devin”x.com
- SOURCE@cognition“scaled RL to multiple trillions of parameters”x.com
- SOURCE@cognition“free for all Pro, Max & Teams subscribers for the next month”x.com
- SUPPORT@cognition“beats SWE-1.7, Grok 4.6, and GPT 5.6 Sol”x.com
- SUPPORT@cognition“first model to support effort levels”x.com
- SUPPORT@cognition“taking 58% fewer turns and costing 81% less on average”x.com