Cognition launches SWE-2 coding model, claiming frontier-level performance at up to 70% lower cost

Cognition launched SWE-2 and made it generally available, with founder Scott Wu announcing it as the company's first model claimed to match frontier models on major benchmarks at up to 70% lower cost. It is available immediately in Devin desktop and CLI, free for one month for all Pro, Max and Teams subscribers. The launch coincided with a $2B funding round and the Devin Voice release, while the marketing drew criticism over cherry-picked benchmarks.

Confirmed

• Official claims: SWE-2 matches recent frontier models on major benchmarks at up to 70% lower cost; Cognition published a blog on how SWE-2 was trained
• Per @carlesgelada: SWE-2 scores 50.0% on FrontierCode 1.1 Main, 1 point below Fable 5.1 but 64% cheaper
• Per @silasalberti's repost, SWE-2 is post-trained on Kimi
• Per @petrusenkomax and @simonguozirui, SWE-2 is claimed to compete with Fable 5.1 and GPT-Astra, marking Cognition's push into foundation coding models after its agent products
• Per @silasalberti, hands-on feedback says SWE-2 outperforms Astra and Fable at lower cost and is not benchmark-optimized; a demo built a website with two custom skills plus Devin CLI, zero prompts, zero cost
• The company raised $2B; per @marvinvonhagen, Devin Voice was released the same day

Technical highlights

• The team scaled RL to multi-trillion parameter scale and improved the RL recipe, pushing the Pareto frontier on both capability and cost
• Per @simonguozirui, training added a penalty term to preserve test-time scaling
• Per @andrewncarr, SWE-2 is Cognition's first model with effort levels, requiring major improvements to the length penalty scheme; a single RL run pushed the Pareto curve. Per @DeryaTR, high-effort performance is especially strong in practice

Unconfirmed

• Per @steipete, critics argue Cognition cherry-picked favorable benchmarks, so the "matches frontier" claim awaits independent verification; all performance and cost claims come from official or supporter accounts, and the "post-trained on Kimi" detail also awaits official confirmation

Why it matters

• SWE-2 offers near-frontier coding capability at much lower prices, potentially intensifying price and performance competition in the coding model market; combined with the $2B raise and rapid-fire releases, Cognition is expanding from agent products into foundation models

2026-09-10 ~ 2026-09-12 · 12 related posts

• Cognition launches SWE-2: frontier-level coding performance at up to 70% lower cost — silasalberti · 2026-09-10
• [source] Cognition Ships SWE-2 at 70% Lower Cost Alongside $2B Raise — DevinAI · 2026-09-11
• Cognition launches SWE-2 coding model, rivaling Fable 5.1 and GPT-Astra — petrusenko_max · 2026-09-11
• Cognition ships SWE-2 in Devin, free for Pro, Max & Teams subscribers for a month — Scobleizer · 2026-09-11
• [source] Cognition Ships SWE-2, Its First Model With Effort Levels, Pushing the Pareto Curve in a Single RL Run — andrew_n_carr · 2026-09-11
• Cognition Launches SWE-2 Model Matching Frontier Evals at 70% Lower Cost, Plus Devin Voice — marvinvonhagen · 2026-09-11
• Cognition launches SWE-2 claiming frontier-level coding at 70% lower cost, but skips Terminal-Bench 4 — steipete · 2026-09-11
• SWE-2 reportedly beats Astra and Fable at lower cost, free for Devin's $20 tier — silasalberti · 2026-09-11
• [source] Cognition ships SWE-2: first model at frontier-level performance, free for Devin users for a month — DeryaTR_ · 2026-09-11
• Cognition launches SWE-2, matching Fable 5.1 on FrontierCode at 64% lower cost — simonguozirui · 2026-09-11

2 near-duplicate retellings: silasalberti · carlesgelada