Cognition, the company behind the Devin coding agent, just shipped SWE-2 Flash — a smaller sibling of its flagship SWE-2 model that runs at roughly 3x the speed for about one-fifth of the price. Announced November 4, 2026, it is available now in Devin Desktop and the Devin CLI, and it is free to use through November 30, 2026.
The launch is interesting for a specific reason: it is one of the most honest model launches of the year. Cognition is not claiming frontier dominance. It is claiming that most daily coding work does not need the frontier — and pricing accordingly.
What it actually is
SWE-2 Flash is a 276-billion-parameter mixture-of-experts model with 12 billion parameters active per token, adapted from Thinking Machines Lab's Inkling-Small base model (weights released July 30, 2026). That MoE structure is the whole trick: sparse activation keeps inference fast and cheap while the total parameter count stays large enough to handle real coding work.
The intended workloads, per Cognition: automations, code review, feature development, and documentation updates. The high-volume, low-drama work that eats most engineering budgets.
Where it wins — and where the big model still wins
Cognition's reported evaluations show Flash delivering big gains over its Inkling-Small base. Against SWE-2 itself, the picture is mixed:
- FrontierCode 2.0: Flash edges out SWE-2 — notable for a model at a fifth of the cost.
- Terminal-Bench 4.0: SWE-2 scores higher. Long, complex terminal tasks still belong to the flagship.
- DeepSWE 1.1: SWE-2 scores higher here too.
One honest caveat to carry with you: Cognition's own speed illustration is a simulation — each model's Terminal-Bench run replayed at 200 vs. 60 tokens per second. It is marketing visualization, not a measured benchmark. Treat all vendor-reported numbers as best-case until independent evaluations land.
The pricing math
Cognition hasn't published a separate per-token table for Flash — pricing is positioned strictly relative to SWE-2. SWE-2's Devin list price is $3.00 per million input tokens and $15.00 per million output tokens, so one-fifth works out to roughly $0.60 input / $3.00 output per million tokens.
The catches: Flash runs only inside Devin (Desktop and CLI) — no standalone API, no public weights. Cognition says it is building a secure sharing program for trusted partners and research labs instead. If you want to self-host, look at open-weight models instead.
The takeaway for developers
The thesis here is the same efficiency story reshaping the whole industry: a mid-size MoE can handle the daily grind. If your agent spend is dominated by routine work, the free window through November 30 is a genuinely free way to test whether the cheap-fast model covers your workflow. Just measure before the December billing reality sets in.
This is a condensed developer explainer. The full breakdown — what changed versus SWE-2, the complete benchmark picture, who should and shouldn't use it, and the try-it-free steps — is on NeurAITool: SWE-2 Flash: Cognition's Cheap, Fast AI Coder (2026).
Top comments (0)