Closed SourceCognitionReleased 2026-09

SWE-2

Frontier coding performance at 64% lower cost

Context window

1M

Input / 1M tokens

Not announced

Output / 1M tokens

Not announced

Provider

Cognition

Data verified 2026-09-12

SWE-2 is Cognition's new coding model, post-trained from Moonshot AI's Kimi K3 (2.8T parameters) using reinforcement learning with configurable reasoning-effort levels. It scores 50.0% on FrontierCode 1.1 Main—within one point of Claude Fable 5.1—while Cognition claims it costs 64% less at that benchmark. Available exclusively through the Devin platform.

Capability index

Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.

Coding
9
Reasoning
6
Math
0
Multimodal
0
Long context
8
Speed
8
Cost efficiency
9

How to access it

Available in Devin Desktop, CLI, Web, and Fusion. No standalone API or open weights. Devin Pro starts at $20/month; one-month free trial for all Pro, Max, and Teams subscribers.

Strengths

  • Strong performance on short, well-scoped coding tasks (92.8% on Terminal-Bench 2.1)
  • Configurable reasoning effort levels (medium, high, max) optimized on the Pareto frontier
  • Cost-efficient: 64% cheaper than Fable 5.1 at equivalent FrontierCode performance
  • Integrated directly into Devin's full software engineering agent

Best for developers who...

Routine PR reviews and small feature implementationsBug fixes and incremental coding tasksTeams already using Devin's ecosystem

When to choose it (and when not to)

Reach for SWE-2 when...

  • You need coding AI at a significant cost discount
  • You are already on the Devin platform
  • You work on short-horizon, well-scoped coding tasks

Look elsewhere if...

  • You need long-horizon agentic reasoning (scores 27.3% on Terminal-Bench 4 vs 57.9% for GPT-6 Astra)
  • You require standalone API access or custom agent harnesses
  • You need to integrate via OpenRouter or custom CI/CD pipelines
  • Your workflow requires open weights

How to use it

  • Set effort level (medium/high/max) based on task complexity and cost tolerance
  • Works best with well-scoped, single-file or multi-file changes within a repository
  • Use within Devin's managed environment for optimal performance

Quickstart

N/A
// No standalone API. Access via Devin Desktop, CLI, or Web at devin.ai

SWE-2 is not available through OpenAI API, Anthropic API, or standard model routers. Use through Devin products only.

API model id: N/A - Devin products only

Benchmarks

BenchmarkScoreNotes
FrontierCode 1.1 Main50.0%Within 1 point of Fable 5.1 (50.9%), 3+ points behind GPT-6 Astra (53.3%)
Terminal-Bench 2.192.8%Short-horizon task performance
Terminal-Bench 427.3%Long-horizon agentic reasoning; significantly weaker than frontier models
DeepSWE 1.1Not disclosedInternal result published

Source: Cognition Official Launch Post

Compare SWE-2

Compare SWE-2 with any other model

Build a comparison →
All model comparisons →

Learn the concepts