Cognition Launches Devin Voice and SWE-2 Coding Model

Cognition introduced Devin Voice, letting a user describe a coding task by phone and have Devin carry it out. The feature is built on OpenAI's GPT-Live-1 for the voice interaction and Cognition's new SWE-2 model, post-trained from Kimi K3, for the coding work, which Cognition describes as reaching frontier-level coding performance with tunable reasoning effort at a fraction of its predecessor SWE-1.7's cost.

That cost and performance claim is Cognition's own account rather than an independent test. A related item in this batch describes Artificial Analysis independently benchmarking a GPT-6 Astra plus SWE-2 pairing inside Devin Fusion, finding it scored close to a Claude Fable 5.1 pairing while costing 43% less.