Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra
36 points by seelos 55 minutes ago | 6 comments
monkeydust 6 minutes ago
As an Econ graduate, pretty cool seeing Pareto in the "AI-bro" zeitgeist. Slightly surreal watching a 1906 welfare economics idea get rediscovered as a plotting convention. The original, if anyone fancies 579 pages of Italian: https://archive.org/details/manualedieconomi00pareuoft. There is an English translation somewhere.
replymydreamof 14 minutes ago
Seems like benchmaxing? For example for Terminal-Bench 4 it doesn't have great results. And why not show other benchmarks?
replyharmonic18374 7 minutes ago
Probably, FrontierCode is made by Cognition itself. The model also seems worse in every way than DeepSeek v4.1 Flash, launched today.
replyAlso the submitter's account is very new which makes me suspicious of self-promotion.
scronkfinkle 11 minutes ago
Please correct me if I'm wrong, but this appears to require Devin to use? I'm disappointed to see I need to use a bespoke platform to interact with this agent, to the point that I probably won't be trying it.
reply_doctor_love 14 minutes ago
SWE-1.5 was surprisingly good when I used it last. I feel like Cognition is one of the solid players that’s flying a bit under the radar while Anthropic and OpenAI race to IPO.
reply