Cognition's SWE-2 Flash: 3x faster, 1/5 the price, free until Nov 30

The new Devin model trades a bit of accuracy for big speed — plus Meta kills Conductor's cloud Nov 20 and Claude Code background agents turn 99% cheaper.

Nowline NOV 5 9:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Free to use today in Devin Desktop and the CLI

    SWE-2 Flash is live now in Devin Desktop and the CLI, free through November 30, and Cognition pitches it for automations, code review, feature work, and doc updates. It's a drop-in way to cut the cost of the routine agent runs you already do, while keeping a bigger model on call for the hard stuff.

  • One-fifth the price, 3x the speed — with a real tradeoff

    Cognition clocks Flash at ~200 tok/s against SWE-2's ~60, at a fifth of the cost, and claims it beats every American open-weight model up to 3x its size (276B total, 12B active). The honest catch: it gives back accuracy — 20.0 vs 27.3 on Terminal-Bench 4.0 and 60.1 vs 73.0 on DeepSWE 1.1 — so route your gnarliest tasks to the full model.

  • Meta bought Conductor — its cloud shuts down Nov 20

    If you run Conductor's multi-agent coding workspaces in the cloud, start migrating: Meta confirmed the acquisition and the hosted service goes dark November 20, after which user cloud data begins getting deleted. Local and other core features keep working into early next year, and founder Charlie Holtz is teasing 'something new.'

  • Claude Code's background agents got ~99% cheaper to run

    Anthropic now has Haiku 5.5 write background-agent progress updates from short notes instead of re-reading the whole conversation, cutting per-update cost by roughly 99%. Net effect: sessions that lean on background agents may burn about 5% less of your usage allowance — free headroom if you live in Claude Code.

  • New paper: your agent's long context is mostly wasted money

    A fresh arXiv paper argues most long-context spend is dead weight: its End-of-Conversation and Single-Action compaction tricks land within -3.7 to +5.6 points of full-history accuracy while cutting cost 6.5–54.6x. Worth testing on your own agent loops, since generating at 256K context runs 3.6–7x pricier than at 16K.

  • OpenAI's security agents found 948 Linux root exploits for $1.2M

    An OpenAI pipeline on an upcoming cyber model scanned ~2.8M lines of the Linux kernel and produced working privilege-escalation exploits for 948 distinct bugs — about $1,279 each in API credits. 61 are already patched upstream, signaling a wave of AI-driven vuln-hunting (and fixes) landing in the dependencies you ship on.