← Retour au blog
tech 10 September 2026

Cognition Launches SWE-2 Model: A New Rival for Fable 5.1 and GPT-Astra

Cognition introduces SWE-2, an advanced coding model that surpasses its predecessors and rivals Fable 5.1 and GPT-Astra in performance and cost.

Article inspired by the original source
Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra ↗ cognition.com

Introduction

The automated coding industry is abuzz with the arrival of SWE-2, Cognition's latest model. This model promises to revolutionize the landscape by pushing the boundaries of performance and cost. Competing with giants such as Fable 5.1 and GPT-Astra, SWE-2 positions itself as an indispensable player.

SWE-2: A Significant Advancement

SWE-2 was designed to outperform its predecessors while being more economical. The model achieves 50.0% on the FrontierCode 1.1 Main benchmark, getting close to Fable 5.1 at a 64% reduced cost. This achievement results from an unprecedented scale-up of reinforcement learning (RL) into a multi-trillion parameter regime.

Scalability and Efficiency

Building on the SWE-1.72 training infrastructure, SWE-2 introduces a new RL algorithm that trains all reasoning-effort levels in a single run. This advances the entire cost-performance frontier. The model, post-trained from Kimi K33, a 2.8T parameter model, benefited from continuous optimization, adding 5 to 6 points on many benchmarks.

Comparison with Competitors

Performance Benchmark

SWE-2 stands out not only for its performance but also for its cost-effectiveness. On the FrontierCode 1.1 Main benchmark, SWE-2 scores 50.0%, surpassing SWE-1.7 and Grok 4.6, and nearing GPT-6 Astra at a quarter of the price. Additionally, on DeepSWE 1.1, SWE-2 reaches 73.0%, positioning itself very close to Fable 5.1.

Cost Optimization

SWE-2's cost approach involves applying linear penalties per effort level, adjusted to reflect actual user costs. This method, derived from first principles, preserves the Pareto frontier's shape while advancing its limits.

Technological Innovations

Reward Baselines and RL Environments

Cognition introduced advancements such as length-weighted reward baselines, significantly stabilizing training. By tripling the number of RL environments and adding instruction-following overlays, SWE-2 enhances the robustness of its results.

Training Efficiency

With the use of NVFP4/FP8 kernels and quantization-aware training, SWE-2 reduces total memory usage, enabling a more optimized training-inference match than SWE-1.7 despite having a base model nearly three times larger.

Conclusion

SWE-2 is not just an advanced coding model; it redefines performance and cost standards in the industry. Its ability to rival models like Fable 5.1 and GPT-Astra makes it a strategic choice for companies looking to optimize their software development processes.

Let's discuss your project in 15 minutes.

SWE-2 Cognition AI models Fable 5.1 GPT-Astra
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call