Cognition unveils new SWE-2 model, rivaling Fable 5.1 and GPT-Astra
gpt-5 grok
| Source: HN | Original article
Cognition has unveiled its SWE-2 model, which competes with leading AI systems such as Fable 5.1 and GPT‑Astra and outperforms its predecessor and Grok 4.6 in benchmark tests.
Cognition has unveiled SWE‑2, its latest autonomous software‑engineering model, positioning it as a direct challenger to leading coding AIs such as Fable 5.1 and OpenAI’s GPT‑Astra. The company says SWE‑2, built on a base model further refined with post‑training on the Kimi‑K3 dataset and an expanded reinforcement‑learning recipe, delivers “frontier‑level” performance at a fraction of the price of its rivals.
On the FrontierCode 1.1 Main and DeepSWE 1.1 benchmarks, SWE‑2 posted a 50.0 % score, outpacing Cognition’s previous SWE‑1.7 and xAI’s Grok 4.6, while matching Fable 5.1’s results with roughly 64 % lower cost. Across the same tests the model also beat GPT 5.6 Sol and came within a few points of GPT‑6 Astra, yet at about a quarter of the expense. Cognition highlights a “70 % lower cost” ceiling on the benchmarks it ran, underscoring the economic advantage it claims to offer developers.
The launch matters because coding‑focused models have become a battleground for AI firms seeking to capture the lucrative software‑automation market. By delivering comparable accuracy at substantially reduced compute spend, SWE‑2 could lower barriers for enterprises and individual developers to adopt AI‑assisted programming, potentially reshaping pricing dynamics and prompting rivals to revisit their cost structures. The announcement also follows Cognition’s recent $48 billion valuation, signalling investor confidence that the company can translate technical gains into market share.
Going forward, observers will watch how quickly SWE‑2 is integrated into real‑world development pipelines, whether third‑party benchmarks confirm Cognition’s claims, and how competitors such as Anthropic, OpenAI and xAI respond with performance or pricing adjustments. The next wave of benchmark releases and customer case studies will be key indicators of whether SWE‑2 can sustain its touted advantage in the fast‑moving AI coding arena.
Sources
Back to AIPULSEN