Cognition SWE-2 Flash: one-fifth the price, mixed benchmarks
Cognition's SWE-2 Flash coding model runs at 3x the speed and one-fifth the price of SWE-2, free in Devin through Nov 30. Test it on your own repo first.
Cognition released SWE-2 Flash, a smaller coding model that it says runs at 3x the speed and one-fifth the price of SWE-2. The announcement went up on November 4, and the model is free in Devin Desktop and the Devin CLI through November 30. Cheaper coding agents are good news. But Cognition's own benchmark table shows Flash behind SWE-2 on two of the three headline evals. The free month is the right time to find out which side of that gap your codebase sits on.
What actually happened
SWE-2 Flash is built on Thinking Machines' Inkling-Small, a mixture-of-experts model with 276B total parameters and 12B active. Cognition post-trained it for coding agent work. Per Cognition's post:
- Price and speed: one-fifth the price and 3x the speed of SWE-2. No per-token prices were published.
- Availability: Devin Desktop and CLI today, free through 11/30. No public API was announced.
- Weights: not released. Cognition says it is building a program to share them with trusted partners, citing cyber risk. Its ExploitBench score rose from 11.3% to 34.8%.
The SWE-2 Flash benchmarks, from Cognition's own table:
| Benchmark | SWE-2 Flash | SWE-2 |
|---|---|---|
| FrontierCode 2.0 | 24.0 | 23.8 |
| Terminal-Bench 4.0 | 20.0 | 27.3 |
| DeepSWE 1.1 | 60.1 | 73.0 |
Flash edges SWE-2 on one eval and trails by 7 to 13 points on the other two. The speed demo in the post is a simulated replay at 200 vs. 60 tokens per second, not a live measurement.
Why it matters for your business
A coding agent's cost is not the token price. It is the token price times the number of retries, plus the human time spent fixing what it got wrong. A model that is 80% cheaper per call but fails more often on multi-step terminal work can cost more per merged pull request. We covered the same trade-off when SWE-2 launched.
The practical play: route by task. Flash-class models are a good fit for well-scoped work — test scaffolding, small refactors, lint fixes, docs. Keep the bigger model for long, multi-file changes and anything touching deploy scripts or infrastructure. Use the free window to measure it:
- Pick 10 to 20 real tickets you already closed. Run them through Flash.
- Track merged-without-edits rate, retries and review time, not just speed.
- Write the result down before December 1, when the price stops being zero.
Key takeaways
- SWE-2 Flash: 276B total / 12B active parameters, built on Thinking Machines' Inkling-Small
- Cognition claims one-fifth the price and 3x the speed of SWE-2; free in Devin through November 30
- Its own table shows it ahead on FrontierCode 2.0 by 0.2 points and behind on Terminal-Bench 4.0 and DeepSWE 1.1
- Weights are withheld for now over cyber-capability concerns
- Measure cost per merged PR on your own tickets, then route small tasks to the cheap model
Know what your coding agents actually cost. We set up model routing and per-task cost tracking so cheap models do the cheap work and nothing ships unreviewed. See how we build or tell us about your stack.
Sources: Cognition — Introducing SWE-2 Flash, Unite.AI.
- #cognition
- #swe-2-flash
- #devin
- #coding-agents
- #ai-costs
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenAI AI pipeline found 948 Linux root exploits for $1.2M
OpenAI's AI pipeline built 948 Linux kernel root exploits in one $1.2M run. An unpatched server with any low-privilege login is now a root box.
Read itMeta acquires Conductor: export your agent workspaces by Nov 20
Meta bought Conductor, the Mac app for parallel coding agents. Conductor Cloud goes read-only November 13 and data deletion starts November 20. Export now.
Read it