Workspaces
Log in Sign up

Cognition Launches SWE-2: Matching Frontier Coding AI at a Fraction of the Cost

Updated

Sources

Cognition Launches SWE-2: Matching Frontier Coding AI at a Fraction of the Cost

On September 10, 2026, Cognition, the startup behind the autonomous coding agent Devin, officially released SWE-2, its most advanced in-house coding model. Post-trained from Moonshot AI's 2.8-trillion-parameter Kimi K3 base model, SWE-2 is designed specifically to optimize the "Pareto frontier" of agentic software engineering.

By utilizing a single reinforcement learning (RL) training run that spans multiple effort levels, Cognition claims SWE-2 can match or come within one point of today's leading frontier models on real-world coding benchmarks while slashing operational costs by up to 70%.1

The Cost Pitch: 64% Cheaper Than Fable 5.1

SWE-2's primary value proposition is its dramatic cost efficiency. On FrontierCode 1.1 Main—a benchmark built around whether a human repository maintainer would actually merge an AI-written pull request—SWE-2 scored 50.0%. This lands it within a single point of Anthropic’s flagship Fable 5.1 (which scored 50.9%), while being 64% cheaper to run.

SWE-2 also scored within a few points of OpenAI’s brand-new flagship GPT-6 Astra (which scored 53.3%) at approximately a quarter of the cost. By comparison, SpaceXAI's Grok 4.6 scored 48.0%, and Cognition’s previous model, SWE-1.7, scored 42.0%.

The technical efficiency translates directly to workflow improvements. On the medium-effort setting of FrontierCode 1.1 Main, SWE-2 beat the overall score of SWE-1.7 while taking 58% fewer turns and costing 81% less on average. Additionally, SWE-2 reached its first valid code edit after a median of 18 steps, down from 48 steps for SWE-1.7, meaning teams spend less time watching the agent wander through irrelevant files.

The Agentic Gap: Terminal-Bench 4 Warnings

While SWE-2 excels at standard, self-contained software engineering tasks, Cognition's own release includes a warning label regarding harder, more complex agentic environments.

On Terminal-Bench 4—a benchmark designed to test highly complex command-line agent tasks—SWE-2's performance dropped sharply:

  • GPT-6 Astra: 57.9%
  • Fable 5.1: 55.8%
  • SWE-2: 27.3%

This massive performance gap indicates that while SWE-2 is highly optimized for clean, structured software engineering lanes, the expensive, general-purpose frontier models from Anthropic and OpenAI still maintain a massive lead on open-ended, highly agentic tasks.2

Deployment and Market Context

SWE-2 has been rolled out directly inside Devin Desktop and the Devin CLI, with Devin Web and Devin Fusion implementations following. Cognition has not released the weights or a standalone public API for SWE-2, choosing instead to keep it as an exclusive, built-in engine for Devin.

The release capped off an exceptionally loud week for Cognition. On September 8, 2026, the company announced a massive Series E funding round of over $2 billion at a $48 billion valuation.

The timing of the launch also capitalizes on growing developer frustration with "AI slop." OpenAI recently faced backlash after current and former Caltech mathematicians criticized the company's planned "Mathathon," warning of AI-generated math slop that falls apart under human verification. By focusing on a benchmark (FrontierCode 1.1) that measures actual human pull-request merges, Cognition is pitching SWE-2 as a reliable, highly verified tool that delivers real-world utility rather than unverified leaderboard claims.

  1. An instance of Low-cost open-weight models have broken the pricing power of elite proprietary APIs. — By post-training on an open-weight base, Cognition created a specialized model that matches frontier capability while breaking expensive proprietary API pricing. ↩︎

  2. An instance of Frontier intelligence is becoming a commodity while agentic workflows remain a proprietary moat. — Cognition's specialized SWE-2 model undercuts premium options on coding efficiency but is completely outclassed on complex agentic workflows where frontier reasoning engines maintain a moat. ↩︎

Part of

This finding is an example of a pattern recurring across your work:

Revision history

  • Create a new finding to document Cognition's release of SWE-2, its 64% cost savings over Fable 5.1, its post-training from Kimi K3, and its performance on FrontierCode 1.1 vs. Terminal-Bench 4.
    · by the agent
  • Create a new finding to document Cognition's release of SWE-2, its 64% cost savings over Fable 5.1, its post-training from Kimi K3, and its performance on FrontierCode 1.1 vs. Terminal-Bench 4.
    · by the agent
  • Create a new finding to document Cognition's release of SWE-2, its 64% cost savings over Fable 5.1, its post-training from Kimi K3, and its performance on FrontierCode 1.1 vs. Terminal-Bench 4.
    · by the agent
  • Create a new finding to document Cognition's release of SWE-2, its 64% cost savings over Fable 5.1, its post-training from Kimi K3, and its performance on FrontierCode 1.1 vs. Terminal-Bench 4.
    · by the agent
  • Create a new finding to document Cognition's release of SWE-2, its 64% cost savings over Fable 5.1, its post-training from Kimi K3, and its performance on FrontierCode 1.1 vs. Terminal-Bench 4.
    · by the agent
  • Create a new finding to document Cognition's release of SWE-2, its 64% cost savings over Fable 5.1, its post-training from Kimi K3, and its performance on FrontierCode 1.1 vs. Terminal-Bench 4.
    · by the agent