Overview
SWE-2 is the coding model behind Cognition's Devin, released on 10 September 2026. Rather than pre-training from scratch, Cognition post-trains an existing open base model: SWE-2 starts from Moonshot AI's Kimi K3, which Cognition describes as a 2.8-trillion-parameter model that had already undergone extensive reinforcement learning for agentic coding. The same recipe produced the previous generation, SWE-1.7, on top of Kimi K2.7.
The headline claim in the launch post is a price-performance one. On FrontierCode 1.1 Main — Cognition's own benchmark, which asks whether a project's maintainer would actually merge the pull request a model wrote — SWE-2 scores 50.0% against 50.9% for Fable 5.1, and Cognition says it does so at 64% lower cost. Against GPT-6 Astra, which leads that table at 53.3%, Cognition describes SWE-2 as coming within a few points at a quarter of the cost. Compared with its own predecessor, SWE-2 medium is reported to score higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average.
The model is tuned for the long-horizon, asynchronous work Devin does rather than for chat: it leads Cognition's published table on DeepSWE 1.1 (73.0%) and Terminal-Bench 2.1 (92.8%), while Terminal-Bench 4 — the hardest row in that table — remains a wide gap, with SWE-2 at 27.3% against 57.9% for GPT-6 Astra. Weights are not published; access is through Cognition's own products.
| Released | 2026-09-10 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Parameters | Inherited from Kimi K3, which Cognition describes as a 2.8T-parameter model |
| Architecture | Post-trained from Moonshot AI's Kimi K3, a model Cognition says had already undergone extensive reinforcement learning for agentic coding |
| Modalities | Text |
| Status | Generally available — Cognition shipped SWE-2 in Devin Desktop and CLI on release day, with a rollout to Devin Web and Fusion announced alongside it. |
Benchmarks
SWE-2 against the field, as published by Cognition at launch
| Benchmark | SWE-2 | Kimi K3 | Grok 4.6 | Fable 5.1 | GPT-5.6 Sol | GPT-6 Astra | SWE-1.7 |
|---|---|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 50% | 44.2% | 48% | 50.9% | 47.5% | 53.3% | 42% |
| DeepSWE 1.1 | 73% | 68.5% | 67.5% | 67.4% | 72.7% | 74.1% | 37.7% |
| Terminal-Bench 2.1 | 92.8% | 88.3% | 88.4% | 91.4% | 88.8% | 89.9% | 81.5% |
| Terminal-Bench 4 | 27.3% | 21.5% | 20.3% | 55.8% | 37.3% | 57.9% | 7.6% |
This model's scores
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Best score in Cognition's published table on Terminal-Bench 2.1 (92.8%) and DeepSWE 1.1 (73.0%)
- Near-parity with Fable 5.1 on FrontierCode 1.1 Main (50.0% vs 50.9%) at what Cognition reports as 64% lower cost
- Large efficiency gain over the previous generation: SWE-2 medium takes 58% fewer turns than SWE-1.7 and costs 81% less on average
- Built on a base model that had already been reinforcement-learned for agentic coding, so post-training targets Devin's long-horizon workflow directly
- Graded on mergeability rather than test-passing alone — FrontierCode tasks are written by the maintainers of the repositories they come from
Best for
- Long-running, asynchronous software engineering tasks delegated to Devin rather than typed turn by turn
- Pull requests that have to survive maintainer review — scope discipline and codebase style, not just a passing test suite
- Terminal-driven agentic work, where Cognition's published Terminal-Bench 2.1 score is the strongest in its table
- Teams weighing coding-model spend, given the published cost comparison against Fable 5.1 and GPT-6 Astra
How to access
| Provider | Model ID |
|---|---|
| Devin (Cognition) ↗ | — |
FAQ
Is SWE-2 open source?
No. Cognition publishes benchmark results and a launch post, but not weights. SWE-2 is reached through Cognition's own products — Devin Desktop and CLI at launch, with Devin Web and Fusion following.
What model is SWE-2 built on?
Moonshot AI's Kimi K3, which Cognition describes as a 2.8-trillion-parameter model that had already undergone extensive reinforcement learning for agentic coding. Cognition applies its own post-training on top. The previous generation, SWE-1.7, used Kimi K2.7 the same way.
How does SWE-2 compare with Fable 5.1 and GPT-6 Astra?
In Cognition's published table, SWE-2 scores 50.0% on FrontierCode 1.1 Main against 50.9% for Fable 5.1 and 53.3% for GPT-6 Astra. Cognition reports the run costs 64% less than Fable 5.1 and a quarter of GPT-6 Astra. SWE-2 leads the same table on Terminal-Bench 2.1 (92.8%) and DeepSWE 1.1 (73.0%), and trails badly on Terminal-Bench 4 (27.3% against 57.9%).
What is FrontierCode?
A benchmark Cognition publishes that measures mergeability — whether a project maintainer would accept the pull request a model produced. Tasks are written by the maintainers of the repositories they come from, and grading combines unit tests, rubrics and additional verifiers across correctness, test quality, scope discipline and codebase style.
What is SWE-2's context window?
Cognition does not state one in the launch post, so we do not list a figure.