# SWE-2

Cognition, released 2026-09-10

**Verdict: conditional.** SWE-2 is Cognition's coding model for Devin, post-trained from Kimi K3, and Cognition claims it lands within one point of Fable 5.1 on FrontierCode 1.1 Main at 64% lower cost. Conditional because every SWE-2 score is vendor-reported, the same table shows it far behind Fable 5.1 and GPT-6 Astra on Terminal-Bench 4, and it runs only inside Devin. Right for Devin Desktop and CLI users; wrong for anyone who needs an API model or independently verified results. Last verified 2026-09-14.

## Pricing

- Input: N/A (no standalone API)
- Output: N/A (no standalone API)
- Context: Devin only; Enterprise $0.75 in/$3.75 out per 1M

## Strengths

- Devin Desktop and Devin CLI users: SWE-2 has been available there since September 10, 2026, with Devin Web and Fusion rolling out
- Terminal-Bench 2.1-style tasks, where Cognition reports 92.8%, ahead of Fable 5.1 (91.4%) and GPT-6 Astra (89.9%) in the same table
- Picking an effort level per task: it ships as SWE-2 Medium, High and Max, which Devin's Enterprise SaaS model list bills at 6x, 9x and 12x credit multipliers, with High marked recommended
- Cost-sensitive agentic coding, if Cognition's claim holds: 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while being 64% cheaper, with costs assuming list pricing including public discounts
- Terminal-Bench 4 workloads: Cognition's own table puts SWE-2 at 27.3%, against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra
- Anyone who needs an API or open weights: MarkTechPost reports no open weights and no standalone API, Devin only
- Buyers who need independent verification: every SWE-2 score comes from Cognition's post, which reports public results where they exist and otherwise evaluates each model in the harness it was primarily developed for
- Planning around context length, since Cognition's post publishes no context window

## Benchmarks

- FrontierCode 1.1 Main (Cognition-reported): 50.0% (https://cognition.com/blog/swe-2)
- DeepSWE 1.1 (Cognition-reported): 73.0% (https://cognition.com/blog/swe-2)
- Terminal-Bench 2.1 (Cognition-reported): 92.8% (https://cognition.com/blog/swe-2)
- Terminal-Bench 4 (Cognition-reported): 27.3% (https://cognition.com/blog/swe-2)

## Sources

- [Cognition SWE-2 announcement](https://cognition.com/blog/swe-2)
- [MarkTechPost, Cognition Releases SWE-2](https://www.marktechpost.com/2026/09/12/cognition-releases-swe-2-a-kimi-k3-post-trained-coding-model-that-matches-fable-5-1-on-frontiercode-at-64-lower-cost/)
- [Devin Desktop models (fetch truncated, SWE-2 rows not returned)](https://docs.devin.ai/desktop/models)

Canonical: https://neomanex.com/models/swe-2