SWE-2

What is SWE-2?

SWE-2 is Cognition's current in-house coding model, post-trained with reinforcement learning from Moonshot's 2.8-trillion-parameter Kimi K3. Cognition trained its Medium, High and Max effort levels together in one run and penalised cost during training, so it wanders less: at medium effort it starts editing far sooner than SWE-1.7 and costs 81% less on average.

It replaced SWE-1.7 as Devin's flagship and runs in Devin Desktop and the Devin CLI, with a research preview in Devin's cloud agent. Cognition lists it only inside Devin, with no separate API or weights, so you cannot bring it to another coding tool.

What changed from SWE-1.7?

It is smarter and much leaner: a bigger base model, three effort levels, and far less wandering before it edits.

Base model

Kimi K3, with 2.8 trillion parameters, instead of Kimi K2.7.

Cost per task

At medium effort, 81% lower on average with 58% fewer turns on FrontierCode 1.1 Main.

Scores

50.0% against 42.0% on FrontierCode 1.1 Main, and 73.0% against 37.7% on DeepSWE 1.1.

Effort levels

Medium, High and Max, trained together in one run, where SWE-1.7 had Medium and Max.

Who is SWE-2 best for?

SWE-2 is best for developers who already work in Devin Desktop or the Devin CLI and want a coding model near frontier scores at a lower cost per task, including as the sidekick in Devin's Fusion pairings.

• Everyday coding in Devin

  Cognition reports 50.0% on FrontierCode 1.1 Main, within a point of Fable 5.1, at 64% lower cost.

• Quick fixes

  At medium effort it makes its first real edit after a median of 18 steps, against 48 for SWE-1.7.

• The Fusion sidekick

  Devin recommends Fable 5.1 to plan and review with SWE-2 doing the implementation, so most of the routine work runs at SWE-2's lower rate.

• Harder multi-file changes

  High and Max effort plan more, read more of the codebase and check their work more carefully.

What's included with SWE-2?

Three effort levels at one price, focused codebase exploration, end-to-end tests, self-checking answers and a place in Devin's Fusion pairings.

Three effort levels

Medium, High and Max, trained together in one run and priced the same per token.

Focused exploration

It judges which files matter, taking 53 steps per FrontierCode task at medium effort against 127 for SWE-1.7.

End-to-end tests

Cognition says it writes tests that exercise the whole change and catch regressions and edge cases more reliably.

Checks its own conclusions

When challenged it works the answer out again and runs code for evidence instead of simply agreeing.

Fusion pairing

In Devin's Fusion mode a lead model such as Fable 5.1 plans while SWE-2 writes code and runs builds and tests.

How much does SWE-2 cost?

SWE-2's list price is $3 per million input tokens and $15 per million output tokens, charged against a paid Devin plan's quota; plans start with Pro at $20 a month, and the Free plan does not include it.

SWE-2 is not on Devin's Free plan, which only runs SWE-1.6. On paid plans SWE-2 draws down the daily and weekly quota at its list price, and usage past the quota is billed at the same price from prepaid credits that never expire. Enterprise is priced on request, with SSO, a dedicated deployment option and a dedicated account team. Devin shows monthly prices only. Full pricing on devin.ai. We checked these prices on 28 Sep 2026.

How do readers rate SWE-2?

No one has rated SWE-2 yet. Be the first.

SWE-2 pros and cons

It scores close to Fable 5.1 on Cognition's benchmark at a far lower cost per task, but it runs only in Devin and trails the frontier on the newer Terminal-Bench 4.

What are the best SWE-2 alternatives?

Four tools on Recaply do SWE-2's job. Laguna S 2.1 is open source, so you can run it on your own server.

Claude Fable 5.1Paid

Choose it when the task is hard: Cognition's own table puts it at 55.8% on Terminal-Bench 4 against 27.3% for SWE-2, though it costs $10 / $50 per million tokens; Devin's Fusion can pair the two.

Claude Opus 5.5Paid

Choose it when you want a model you can use in any tool: it runs on the Claude API, AWS, Google Cloud and Microsoft Azure as well as in Devin, at $4 / $20 per million tokens.

Composer 2.5Free

Choose it when you code in Cursor instead of Devin; it is Cursor's own agentic model with a 200K context, at $0.50 / $2.50 per million tokens.

Laguna S 2.1Open Source

Choose it when you want an open coding model you can run in any agent, rather than one locked inside Devin.

Questions about SWE-2