Cognition has launched SWE-2 Flash, a coding model available today in Devin Desktop and CLI. The company says it runs three times faster and costs one-fifth as much as its larger SWE-2 sibling, putting a more affordable model behind everyday software-agent tasks.
Cognition Watch analysis
What happened
Cognition says SWE-2 Flash generates 200 tokens per second, compared with 60 for SWE-2, and is available free through 30 November. The company says the model was trained with a new mid-training stage before reinforcement learning, using a mix that included scientific articles, blogs and code.
Cognition reports that SWE-2 Flash improved its Terminal-Bench 4.0 score from 0.9% to 20.0% and its FrontierCode 2.0 score from 3.2% to 24.0% against the base model. It also reports a 34.8% ExploitBench score, up from 11.3%. The company says the model is not being released with public weights: it is developing a programme for sharing them with trusted industry partners and research labs.
Why it matters
The headline is not simply a new leaderboard entrant. SWE-2 Flash is positioned as a quicker, cheaper option for the coding work software agents already tackle, from automations and code review to feature development and documentation. That could make agent-assisted engineering more practical for teams that do not need the largest model on every task.
The security angle makes the release more consequential. Cognition says it trained for cybersecurity work, while also warning that the same capabilities could be misused. Its answer for now is restricted access to model weights, rather than a public download.
Our read
A model that is both faster and cheaper has a rather better chance of earning a place in the daily workflow than one that merely dazzles on a chart. Cognition’s benchmark gains are promising, but they are company-reported results, not a guarantee that every team will see the same improvement. The free trial gives users a straightforward way to judge the practical difference for themselves.
What to watch
- Whether developers find the speed and cost trade-off useful on real coding tasks.
- How the model performs beyond the benchmarks Cognition highlighted.
- What safeguards and access criteria Cognition sets for its planned weights-sharing programme.
Discussion spark: Would you choose a smaller coding model that is cheaper and faster for routine engineering work, or keep those tasks on the strongest model available?
Sources and evidence
- Source update (4 November 2026, 18:00 UTC)
not affiliated with or endorsed by Cognition