Cognition · 2026-09-10 · major
SWE-2 — Cognition's coding model lands within a point of Fable 5.1
SWE-2 is Cognition's new coding model, post-trained from Moonshot's 2.8-trillion-parameter Kimi K3. It scores 50.0% on FrontierCode 1.1 Main against 50.9% for Fable 5.1, and Cognition says it costs 64% less to run at that score.
SWE-2 post-trains Moonshot's Kimi K3 with reinforcement learning and lands near frontier coding scores for a fraction of the price.
Key specs
| Cost vs fable 5.1 | 64% cheaper |
|---|
Quick facts
| Maker | Cognition |
|---|---|
| Base model | Kimi K3 (2.8T parameters, Moonshot AI) |
| Availability | Devin Desktop and CLI |
| Rolling out to | Devin Web and Fusion |
| What's new | Reasoning-effort levels trained in a single RL run |
| vs SWE-1.7 | 58% fewer turns, 81% lower average cost |
Benchmarks
| GPT-6 Astra | 53.3% | |
|---|---|---|
| Fable 5.1 | 50.9% | |
| SWE-2 | 50% | |
| Grok 4.6 | 48% | |
| GPT-5.6 Sol | 47.5% | |
| Kimi K3 | 44.2% | |
| SWE-1.7 | 42% |
| GPT-6 Astra | 74.1% | |
|---|---|---|
| SWE-2 | 73% | |
| GPT-5.6 Sol | 72.7% | |
| Kimi K3 | 68.5% | |
| Grok 4.6 | 67.5% | |
| Fable 5.1 | 67.4% | |
| SWE-1.7 | 37.7% |
| SWE-2 | 92.8% | |
|---|---|---|
| Fable 5.1 | 91.4% | |
| GPT-6 Astra | 89.9% | |
| GPT-5.6 Sol | 88.8% | |
| Grok 4.6 | 88.4% | |
| Kimi K3 | 88.3% | |
| SWE-1.7 | 81.5% |
What is it?
Configurable reasoning-effort levels are the headline change in SWE-2, Cognition's new coding model, and they are all trained inside one reinforcement-learning run. The model is post-trained from Kimi K3, Moonshot AI's 2.8-trillion-parameter open model, which had already been through heavy RL for agentic coding. Cognition writes that its own RL "still finds substantial headroom, adding 5-6 points on many benchmarks".
How does it work?
The training trick is a cost penalty applied per effort level: "We apply a linear cost penalty per effort level in a single RL run, with each penalty tuned to the local slope of the base model's Pareto frontier." Because every effort level moves together, the whole cost-performance curve shifts rather than one point on it. A user then picks the effort level that fits the budget for a given task.
Why does it matter?
Teams running Devin pay frontier prices for agentic coding today. On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average. Cognition also says the model "comes within a few points of GPT-6 Astra at a quarter of the cost", which changes what a long agent run costs to finish.
Who is it for?
teams running Devin on production codebases
Frequently asked questions
- How much does SWE-2 cost per token?
- Cognition has not published a per-token price for SWE-2. The announcement states cost only in relative terms: 64% cheaper than Fable 5.1 at the same FrontierCode 1.1 Main score, a quarter of GPT-6 Astra's cost for a result a few points behind it, and 81% lower average cost than SWE-1.7 on the same benchmark.
- Where can I use SWE-2 today?
- SWE-2 is available in Devin Desktop and the Devin CLI starting on release day, September 10, 2026. Cognition says it is also rolling the model out on Devin Web and Devin Fusion. The announcement does not mention open weights or a standalone API endpoint for SWE-2 outside the Devin products.
- How much better is SWE-2 than the Kimi K3 model it is built on?
- SWE-2 beats its Kimi K3 base on every benchmark Cognition published: 50.0% against 44.2% on FrontierCode 1.1 Main, 73.0% against 68.5% on DeepSWE 1.1, 92.8% against 88.3% on Terminal-Bench 2.1, and 27.3% against 21.5% on Terminal-Bench 4. Cognition describes the gain as 5-6 points on many benchmarks.
- What is FrontierCode 1.1 Main?
- FrontierCode is Cognition's own benchmark, so the headline number comes from the model's maker. It measures mergeability — whether a repository maintainer would actually merge the pull request an agent wrote. Open-source maintainers write the tasks and define what mergeable means, grading combines unit tests, rubrics and verifiers, and internet access is blocked so models cannot look up solutions.
- Where does SWE-2 still fall behind the frontier models?
- Terminal-Bench 4 is the clearest gap in Cognition's table. SWE-2 scores 27.3% there, against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra, though it still leads Kimi K3 at 21.5% and Grok 4.6 at 20.3%. On FrontierCode 1.1 Main, GPT-6 Astra also stays ahead at 53.3% versus SWE-2's 50.0%.
Try it
curl -fsSL https://cli.devin.ai/install.sh | bash