Anthropicが新しいAIモデル「Claude Opus 5.5」を発表しました。従来のClaude Opus 5と比べて、一般的な処理で40％低いコストでの利用が可能になり、出力速度は30％以上高速化しています。複数の手順を自律的に進める「エージェント型」の作業で性能が向上しており、コマンドライン上での複雑な作業を測る「Terminal-Bench 4.0」では66.4％を記録し、Opus 5の52.3％やGPT-6 Astraの57.9％を上回っています。ソフトウェア開発能力を測る「FrontierCode v1.1」では54.4％、「Curs…

Cognition presenta SWE-2, su modelo de codificación más capaz: 50% en FrontierCode, 64% más barato que Fable 5.1, pero solo disponible dentro de Devin.

CognitionのコーディングモデルSWE-2を早速検証。Kimi K3をポストトレーニングした結果、FrontierCodeでFable 5.1並みの精度を叩き出しつつ、コストを64%も削減できて驚いた。実務のコード生成コスト改善に直…

SWE-2 소개: 코딩 모델의 경쟁 축이 점수에서 비용 대비 성능으로 Cognition이 현지 시간 2026년 9월 10일 자사의 코딩 특화 모델 SWE-2 를 공개했습니다. 핵심 수치는 FrontierCode 1.1 Main에서 50.0%입니다. 같은 벤치마크에서 Claude Fable 5.1이 기록한 50.9%에 1점이 채 못 미치지만, 롤아웃 한 번에 드는 평균 비용은 64% 낮습니다. Cognition은 이 결과를 두고 파레토 프론티어(Pareto Frontier), 즉 더 싸게 만들려면 성능을 포기해야 하는 경계선 자체를 밀어냈다고 표현합니다. C...

Four Signals — The Wire

Cognition's new SWE-2 coding model scores within a point of Fable 5.1 on the FrontierCode 1.1 Main benchmark while costing 64% less, and comes within a few points of GPT-6 Astra at a quarter of its price.

Cognition just launched SWE-2, an AI coding agent it says scores within a single point of Anthropic's top coding model while costing up to 70% less to run. The pitch is blunt: near-frontier code, at a fraction of the bill. Cognition, the startup behind the autonomous coding agent Devin, shipped SWE-2 on September 10, 2026, rolling it out across Devin Desktop, the Devin CLI, Devin Web and Devin Fusion. On the company's own FrontierCode 1.1 Main benchmark, which grades whether a human maintainer would actually merge an AI-written pull request, SWE-2 scored 50.0%. Anthropic's Fable 5.1 scored 50.9%. That's a single point apart. Cognition says it matched that score for 64% less compute cost. Against OpenAI's GPT-6 Astra, which scored 53.3% on the same test, Cognition claims SWE-2 comes in at roughly a quarter of the price. Price is the argument. It also beat its own predecessor, SWE-1.7, which scored 42.0%, and xAI's Grok 4.6, which scored 48.0%. The engineering behind that claim is specific. SWE-2 is post-trained from Moonshot AI's Kimi K3, a 2.8 trillion parameter base model, using a reinforcement learning method that trains three separate effort levels, medium, high and max, in a single run instead of...

2.8조 매개변수 Kimi K3 를 후속 학습한 SWE-2는 FrontierCode 1.1 Main에서 50.0%를 기록해 Fable 5.1과 0.9%p 차이를 보이면서도 비용은 64% 낮음 추론 노력 수준별로 기반 모델의 비용·성능 곡선 기울기에 맞춘 선형 비용 페널티 를 적용해, 한 ...