Cognition SWE-2 is a coding model built for agentic software-engineering work inside Devin. Cognition reports a 50.0% score on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1. Those are Cognition-run results, not independent tests, and harness details matter.
The practical limitation is just as important as the benchmark table. SWE-2 launched in Devin Desktop and Devin CLI, with rollout to Devin Web and Fusion. Cognition has not published a standalone API, token price, context-window specification or downloadable weights. You choose SWE-2 as part of the Devin product rather than integrating it like an OpenAI or Anthropic API model.
SWE-2 at a glance
| Item | Current information |
|---|---|
| Developer | Cognition |
| Release date | September 10, 2026 |
| Primary use | Agentic coding and software-engineering tasks |
| Base model | Kimi K3, according to Cognition |
| Launch access | Devin Desktop and Devin CLI |
| Additional rollout | Devin Web and Fusion |
| Standalone API | Not published |
| Token pricing | Not published |
| Context window | Not published by Cognition for SWE-2 |
| Model weights | Not published |
Do not infer SWE-2’s context limit from Kimi K3 simply because Cognition says it used Kimi K3 as the base. Post-training lineage does not guarantee identical serving limits or product behavior.
What Cognition changed
Cognition says SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter mixture-of-experts model. The company applied reinforcement learning for agentic coding and trained multiple reasoning-effort levels in one run. Its stated aim was not just a higher maximum score, but a better capability-to-cost curve across effort settings.
That matters inside a coding agent. A model that reaches a solution in fewer turns can consume less wall-clock time and fewer tool calls even when its raw token price is not the lowest. Cognition reports that SWE-2 at medium effort uses 58% fewer turns and costs 81% less than SWE-1.7 on FrontierCode 1.1 Main. Treat both figures as vendor claims tied to Cognition’s evaluation setup.
The model also inherits a product context that a bare benchmark cannot represent. Devin provides the repository, shell, browser, editor, planning loop and execution environment. A result achieved inside that harness does not tell you how an unavailable standalone SWE-2 endpoint would behave in another agent.
Cognition’s published benchmark results
| Benchmark | SWE-2 | GPT-6 Astra | GPT-5.6 Sol | Fable 5.1 | Grok 4.6 |
|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 53.3% | 47.5% | 50.9% | 48.0% |
| DeepSWE 1.1 | 73.0% | 74.1% | 72.7% | 67.4% | 67.5% |
| Terminal-Bench 2.1 | 92.8% | 89.9% | 88.8% | 91.4% | 88.4% |
| Terminal-Bench 4 | 27.3% | 57.9% | 37.3% | 55.8% | 20.3% |
These values come from Cognition’s launch post. Cognition notes that some evaluations use public results while others use internal Devin CLI runs at the strongest reasoning setting. The table is useful for understanding the vendor’s positioning, but it is not a neutral leaderboard or proof that SWE-2 will beat another model on your repository.
The mixed result is more useful than a blanket winner claim. SWE-2 is close to the top group on FrontierCode and DeepSWE, leads this vendor table on Terminal-Bench 2.1, and trails Astra and Fable 5.1 substantially on Terminal-Bench 4. That variation is a reminder to evaluate the tasks and harness that match your work.
For a broader explanation of why coding leaderboards disagree, read our AI model leaderboard guide. For current OpenAI positioning, see GPT-6 Astra explained and the GPT-6 Astra vs GPT-5.6 Sol comparison.
What does SWE-2 cost?
Cognition has not published a standalone per-token price for SWE-2. Its launch claims compare end-to-end task cost inside the evaluated Devin environment. That is different from an API price and should not be converted into a fictional input/output rate.
If you are comparing SWE-2 with Fable 5.1, Sol or Astra, compare the customer product you can actually buy:
- Check whether SWE-2 is enabled in your Devin surface and plan.
- Measure the credits or product usage consumed by a representative task.
- Include failed attempts, review time and manual corrections.
- Compare completed, accepted tasks rather than tokens alone.
The relevant unit is cost per accepted software change. A cheaper run that produces an unsafe patch is not cheaper after review and rework.
Where SWE-2 fits
SWE-2 is most relevant when you already use Devin or want an agent that can operate across a repository, terminal and browser. Good evaluation tasks include dependency upgrades, multi-file bug fixes, migrations with test coverage and bounded feature work with clear acceptance criteria.
It is a weaker fit if you need a model endpoint for your own product, require self-hosting, or need published context and pricing limits before procurement. Cognition has not announced those forms of access for SWE-2.
If you want an interactive terminal agent rather than a delegated Devin workflow, compare the broader tool category in our best AI coding tools guide. Our coding agents guide explains the control boundaries that still belong to the developer.
How to evaluate SWE-2 safely
Do not switch a production workflow because of one launch benchmark. Build a small evaluation set from real repository work:
- one well-scoped bug with a failing test;
- one multi-file refactor with behavior-preserving checks;
- one dependency or framework migration;
- one task requiring documentation lookup;
- one ambiguous request that should trigger clarification;
- one security-sensitive change that requires explicit approval.
Record task completion, test pass rate, unwanted file changes, review corrections, total turns and product cost. Use the same repository state and acceptance criteria for every model. The AI testing and evaluation hub covers evaluation datasets and release gates, while the AI agent reliability guide covers tool failures and recovery.
Security and workflow boundaries
A stronger coding model does not remove the need for scoped credentials, isolated environments and review. Keep production secrets out of agent sandboxes, restrict network and deployment permissions, and require approval before external side effects. Review generated migrations and infrastructure changes as carefully as human-authored ones.
Cognition’s benchmark scores measure task performance, not the security of your repository or deployment process. Use the AI application security guide and production agent deployment checklist to build controls around the model.
My take
SWE-2 looks like a serious upgrade for Devin users. Cognition’s published results place it near the frontier on several agentic coding evaluations, and the cost-efficiency story is plausible when fewer turns lead to an accepted change.
But it is not a general replacement for API coding models today. There is no published standalone endpoint, token price, context specification or open-weight release. Test it as a Devin capability, using your own repository and end-to-end task cost. Do not compare an internal Devin run directly with an API token table and call the cheaper one solved.
Frequently asked questions
Is Cognition SWE-2 available through an API?
Cognition has not published a standalone SWE-2 API. At launch, it is available in Devin Desktop and CLI, with rollout to Devin Web and Fusion.
How much does SWE-2 cost?
Cognition has not published per-token SWE-2 pricing. Its launch article reports relative task-cost results inside its evaluation harness. Check current Devin plan and usage terms, then measure cost per accepted task.
Is SWE-2 better than GPT-6 Astra?
Not universally. In Cognition’s table, Astra scores higher on FrontierCode 1.1 Main, DeepSWE 1.1 and Terminal-Bench 4, while SWE-2 scores higher on Terminal-Bench 2.1. Your repository evaluation matters more than one aggregate table.
Is SWE-2 based on Kimi K3?
Yes. Cognition says SWE-2 was post-trained from Kimi K3 using reinforcement learning for agentic coding. Cognition has not said that SWE-2 shares Kimi K3’s public serving limits or licensing.
Can I run SWE-2 locally?
No local weights or self-hosting route have been published. SWE-2 is currently a model offered through Devin product surfaces.