SWE-2 for Coding Agents: What to Know

Published September 17, 2026 · by the AQ team

Cognition announced SWE-2 on September 10, 2026, calling it its most advanced coding model yet: post-trained from Moonshot AI's Kimi K3, and, on Cognition's own evaluations, matching Claude Fable 5.1 on FrontierCode at 64 percent lower cost while posting a 92.8 on Terminal-Bench 2.1. It is available now in Devin Desktop and Devin CLI, free for Devin Pro, Max and Teams subscribers through October 10, 2026, with rollout to Devin Web and Fusion underway. The launch also reopened a familiar argument about what a cross-vendor cost-per-task table really measures, since each model was scored inside its own harness. This page covers what changed, what the benchmark table does and does not show, and how a team actually runs it, as of September 2026.

What SWE-2 actually is

SWE-2 is not a from-scratch foundation model. Cognition post-trained it from Kimi K3, the 2.8 trillion parameter mixture-of-experts model Moonshot AI released with open weights in July 2026, the largest open-weights model shipped to date. Cognition notes that K3 had already been through extensive reinforcement learning for agentic coding, and says its own RL recipe still found substantial headroom on top: around 5 to 6 points on many benchmarks, enough to shift the base model's whole cost-and-performance frontier. That continues the SWE line's pattern (SWE-1.5, 1.6 and 1.7 were also post-trains of strong open bases) and it is the first time Cognition has applied the recipe to a base this large.

The genuinely new piece is effort levels. Cognition describes SWE-2 as the first SWE model with reasoning effort levels, and the training method is the interesting part: rather than training one model and distilling cheaper variants, a single RL run trains every effort level at once, with each level carrying its own linear cost penalty tuned to the local slope of the base model's Pareto frontier. The intent, in Cognition's words, is to move the entire cost-versus-capability curve rather than one point on it, and to make the cost a user actually pays part of the training signal. In practice you pick an effort level the way you pick a reasoning budget on other frontier models, and Cognition's headline table leans on the medium setting.

The numbers Cognition published

All of the launch numbers below are Cognition's own, from its September 10, 2026 announcement, and none have independent replication yet. With that stated, the table Cognition leads with:

FrontierCode is Cognition's own benchmark: real issues from 36 open-source repositories, graded by maintainer-authored rubrics for mergeability (correctness, regression safety, scope discipline, project conventions), not just passing tests, with the 1.1 Main subset holding the 100 hardest tasks. On it, Cognition reports that SWE-2 medium beats SWE-1.7 while taking 58 percent fewer turns and costing 81 percent less on average, and that on FrontierCode 1.1 Main and DeepSWE 1.1 together SWE-2 matches GPT-5.6 Sol and the Fable 5 family at a fraction of their price, coming within a few points of GPT-6 Astra at roughly a quarter of the cost. The headline claim, in Cognition's phrasing on launch day: SWE-2 matches Fable 5.1 on FrontierCode at 64 percent lower cost, with a broader claim of up to 70 percent lower cost on leading evals.

What the table does not show

Three caveats matter, and the first is the one the Hacker News launch thread spent most of its energy on. Cognition's evaluation methodology runs each model in its own harness: Claude Code for Anthropic models, Codex for OpenAI models, and Devin CLI for SWE-2 and other models. That is a defensible choice (it shows each model at its best), but it means every score and every cost figure is a harness-plus-model result, not a model result. A cheaper harness that takes fewer turns lowers cost-per-task regardless of what the weights can do, so the 64 percent figure compares Cognition's full stack against competitors' full stacks at list prices. Our guide to benchmarking the model or the harness covers why this distinction keeps deciding launch-day arguments.

Second, the cost denominator is asymmetric. The comparison uses competitors' published per-task API costs, but Cognition has published no standalone API price for SWE-2 itself as of September 2026: the model is reached through Devin subscriptions, so you cannot recompute the 64 percent from a price sheet. Third, the weak row is real. Terminal-Bench 4, the newest and hardest of the terminal benchmarks, has SWE-2 roughly 30 points behind the frontier, and Cognition's table shows it. The honest reading matches the release name: a Pareto-frontier play (near-frontier scores at a much lower operating point), not a new state of the art, with the newest benchmarks showing where the post-training has not caught up.

Running SWE-2 today

SWE-2 ships where Devin ships. In Devin CLI, install the binary, sign in with your Devin account, and pick SWE-2 with the model flag or the /model command inside a session; Devin Desktop exposes the same choice in its model picker. Through October 10, 2026, SWE-2 usage is free for Pro, Max and Teams subscribers (Devin Pro is 20 dollars a month as of September 2026), which makes this an unusually cheap evaluation window: for a month you can run the same backlog items through SWE-2 and your incumbent model and compare diffs without watching a meter. Cognition says rollout to Devin Web and Fusion is underway; there are no open weights and no standalone API as of September 2026, so Devin surfaces are the only way to run it.

Evaluating it like a team, not a leaderboard

The launch-day numbers cannot tell you whether SWE-2 is right for your codebase, precisely because of the harness caveat above: you would be adopting Devin CLI or Desktop along with the model, and your repository, test suite and conventions are the benchmark that matters. The useful evaluation is the boring one. Pick three or four real issues from your backlog, run each through SWE-2 and through whatever you use today, on the same starting commit, and judge the diffs the way FrontierCode's rubric does: does it merge, does it regress anything, did it stay in scope. Do it before October 10 and the SWE-2 side costs nothing beyond the subscription. Two practical notes from the CLI's design: effort levels are part of the price-performance story, so evaluate the level you would actually run, and long tasks are normal (Cognition's own evaluations allow four hours), so run them somewhere a closed laptop will not kill the session.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. Devin CLI is a first-class harness in AQ: you connect your own Devin account (AQ never marks up usage on your own subscription), and the CLI runs in a persistent tmux session on your team's VM, in an isolated git worktree per workspace, streamed live to the browser. See Devin CLI in AQ for how the connection works, and the launch announcement for what shipped.

That setup happens to be exactly the shape of the evaluation this page recommends. Open two workspaces on the same repository, run SWE-2 through Devin CLI in one and your incumbent through Claude Code or Codex in the other, and let the whole team watch both live sessions rather than reading a summary afterward. Sessions survive a closed laptop and resume from any device, so a four-hour SWE-2 run finishes whether or not anyone is watching, and the free window through October 10 means the experiment costs a Devin subscription and an afternoon.