Cognition's SWE-2 Flash cuts Devin's coding costs to a fifth
The new model runs inside Devin Desktop and the Devin CLI at roughly three times the speed and a fifth of the price of SWE-2, free until 30 November.
Cognition shipped SWE-2 Flash on 4 November, a smaller coding model for its Devin agent that runs inside Devin Desktop and the Devin CLI. Cognition puts it at roughly three times the speed of SWE-2, 200 tokens a second against 60, for about a fifth of the price. It’s free to use until 30 November.
Flash is built on Thinking Machines’ Inkling-Small base, 276 billion total parameters with 12 billion active, against SWE-2’s 2.8 trillion total and 104 billion active. Cognition added a mid-training stage before reinforcement learning, which it says took the base model’s Terminal-Bench 2.1 score from 0.9% to 89.2%. There’s no public API and no released weights: it only runs through Devin, though Cognition says it’s building a programme to share the weights with trusted partners and research labs.
Why it matters: if you’re running Devin for routine changes rather than hard problems, Flash gives you a cheaper, faster default to switch to, and lets you save the full SWE-2 for the tasks that actually need it.
The caveat: the fifth-of-the-price figure is Cognition’s own framing against its own prior model, and it only holds once the free period ends on 30 November, with no confirmed post-promotion rate published yet.