共有:

Coding Agent · Model Update

Devin's model jumps to SWE-2

Two weeks ago SWE-1.7 was still the newest build, with reasoning depth fixed. On September 10, Cognition's own model was replaced by SWE-2 — the first SWE model with selectable reasoning effort. It now scores 92.8% on Terminal-Bench 2.1, and Pro/Max/Teams subscribers get it free through October 10.

AI Navigate Editorial2026.09.207 min read

SWE-1.7 –09.09 fixed effort 09.10 SWE-2 Medium High Max selectable effort 92.8% Terminal- Bench 2.1
01
The Release

SWE-2 shipped into Devin on Sept 10

Cognition Labs turned over the generation of its in-house coding model.

On September 10, 2026, Cognition announced its new model "SWE-2," replacing SWE-1.7 as the newest build. Cognition calls it its "closest model yet to the frontier," and it shipped the same day into Devin Desktop and Devin CLI, with rollout to Devin Web and its multi-model Fusion workflow underway. Details are documented in Cognition's official blog post, "Introducing SWE-2: Pushing the Pareto Frontier", and covered by MarkTechPost.

SWE-2 is post-trained on top of Moonshot AI's 2.8-trillion-parameter "Kimi K3" model. The headline feature is that it's the first SWE model to offer selectable reasoning effort. Previously, the only way to trade off speed against depth was to pick a different model entirely. SWE-2 instead optimizes several cost-performance operating points inside a single reinforcement-learning run, so Cognition no longer has to train and maintain a separate model per effort level — the cost penalty is baked directly into training, which Cognition frames as the key technical departure from earlier SWE models.

02
Numbers

Benchmarks and cost

The key figures that have been published.

92.8%
Terminal-Bench 2.1
50.0%
FrontierCode
-81%
cost vs. SWE-1.7 (Medium effort)

On FrontierCode, Cognition claims parity with Anthropic's "Fable 5.1" at 64% lower cost by its own accounting. But on Terminal-Bench 4, SWE-2 lands at just 27.3%, versus 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra — a reminder that it isn't a universal upgrade.

03
How It Works

Three effort levels

Pick your own cost/accuracy trade-off depending on how heavy the task is.

01

Medium

Jumps into action quickly. Built for simple-to-intermediate tasks, cutting cost by up to 81% versus SWE-1.7 — a fit for routine fixes and small feature additions.

02

High

More planning and codebase exploration, better suited to ambiguous requirements. A balanced pick for mid-size implementation work.

03

Max

Raises the ceiling on turns and tokens, aimed squarely at the hardest tasks with long, complex verification chains.


Optimize the entire
cost-performance curvein a single RL run.


04
Who It Hits

What it means for engineers

The first to benefit are engineers already running Devin on day-to-day work. Choosing Medium effort can cut costs by up to 81% versus SWE-1.7. The more routine fixes and small PRs you already delegate to Devin, the more that cost shift will show up in practice — a genuine reduction in operating cost, not just a benchmark number.

For anyone assigning Devin more complex requirements work or architectural judgment calls, High or Max are worth testing actively. Cases that used to plateau under a fixed reasoning depth may now improve just by dialing the level up. That said, Terminal-Bench 4 shows SWE-2 still trails Fable 5.1 and GPT-6 Astra on tasks with complex verification chains, so maxing out the effort level isn't a cure-all — matching the level to the task still matters.

There's also a practical, time-boxed opportunity here. Pro, Max, and Teams plans get SWE-2 free through October 10. That month is a natural window to test which effort level actually earns its keep on your workload before deciding whether it's worth paying for at full price.

05
Before / After

What changed from SWE-1.7

SWE-1.7 (through 2026.09.09)SWE-2 (2026.09.10–)
Fixed reasoning depthChoose Medium / High / Max
Needed a separate model per levelMultiple performance tiers optimized in one RL run
Base model undisclosedPost-trained on Kimi K3 (2.8T parameters)
No published Terminal-Bench 2.1 score92.8%