SWE-2 Flash: Cognition ships a 3x-faster coding model, free in Devin

276B/12B-active, one-fifth the price of SWE-2, free to try through Nov 30 — as OpenAI's AI finds 948 Linux root exploits and Images 2.6 opens its API.

Nowline NOV 5 1:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A 276B/12B-active coding model, free in Devin through Nov 30

    Cognition's SWE-2 Flash is post-trained from Thinking Machines' Inkling-Small and runs at a simulated 200 tok/s against SWE-2's 60 — roughly 3x the speed at one-fifth the price of its flagship. It's live now in Devin Desktop and CLI, free until Nov 30, with no public API or weights yet.

  • Read the benchmark table before you switch

    Flash doesn't beat the flagship — Terminal-Bench 4.0 20.0 vs 27.3, DeepSWE 1.1 60.1 vs 73.0, SWE-Atlas QA 50.0 vs 56.3 — it only edges SWE-2 on FrontierCode 2.0 (24.0 vs 23.8). Treat it as a cheap, fast default for routine work, not a frontier replacement; a weights-sharing program for partners is 'a few weeks' out.

  • OpenAI's AI found 948 paths to root in Linux

    An unreleased OpenAI cyber model turned 2,308 candidate kernel bugs into 948 working local-privilege-escalation exploits on a stock Linux kernel, for about $1.2M in API spend (~$1,279 per unique bug) in roughly 10 days. Fixes ship through a coordinated 'Patch the Planet' disclosure — update your kernels.

  • Update: Images 2.6 adds two API models, including SVG export

    Beyond the new GIFs in ChatGPT and Codex, OpenAI exposed GPT-Image-2.6 Flare (~6s per image) and Sunburst (native SVG export and image markup) to the API, with generation about 50% faster. SVG-out means editable vector assets straight from a prompt instead of flattened PNGs.

  • Ai2's StateBench breaks top models on long context

    Ai2 released StateBench, a four-tier test of whether a model can track evolving state across long inputs; on the hardest tier leading frontier models fall to 13% or lower. If your agent 'forgets' decisions over a long run, there's finally a benchmark to measure it.