AI developer tools · What shipped

Cognition released SWE-2, an advanced coding model achieving 50.0% on FrontierCode 1.1 Main—within one point of Fable 5.1 while being 64% cheaper. The model was trained using reinforcement learning scaled to the multi-trillion-parameter regime, introducing a novel RL algorithm that trains all reasoning-effort levels in a single run. On multiple benchmarks, SWE-2 outperforms SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and achieves within a few points of GPT-6 Astra at one-quarter the cost. Key technical innovations include a Pareto-informed cost penalty derived from first principles to advance the entire cost-performance frontier, a length-weighted reward baseline that stabilizes training without extra computation, improved RL rollout serving through speculative decoding and online draft model training, and data improvements including tripling RL environments and hardening verifiers. The model demonstrates behavioral improvements in test coverage, resourcefulness, and verification discipline, and is available in Devin Desktop, CLI, Web, and Fusion.

GPT-6 Astra set new performance records on the Epoch Capabilities Index (ECI) as well as math, continual learning, and game-puzzle benchmarks in September 2026. Epoch AI also launched FrontierMath Erdős, featuring 68 unsolved Erdős problems where AI systems write solutions in Lean; GPT-6 Astra solved 2 of 68 problems (3%), the first model to solve any of them. The Epoch AI benchmarking hub tracks 391 models across 85 benchmarks including math, coding, and knowledge reasoning tasks, with evaluation code and detailed model outputs available for transparency and auditability.

Claude Fable 5.1 leads Cursor's CursorBench 4.0 benchmark for multi-file agentic coding tasks with a score of 51.8%, followed by Claude Opus 5 at 46.6% and GPT-5.6 Sol at 41.7%. The benchmark, released September 10, 2026, evaluates long-horizon edit, refactor, investigation, and design-adherence tasks drawn from real Cursor sessions, with 10 models currently evaluated. CursorBench 4.0 is displayed as reference-only on BenchLM and carries a 20% weight in the coding category, though it does not directly affect overall model rankings.

# Extract and Summary Claude Code released v2.1.276 with a critical bug fix for requests failing with HTTP 400 errors when ANTHROPIC_BASE_URL points to a proxy or gateway (regression from v2.1.275). The latest versions include numerous fixes and improvements to developer experience: v2.1.275 added signed-in account confirmation for Claude apps gateway, new send-now key (ctrl+enter) for interrupting turns, and syncing of skills and plugins from claude.ai accounts to terminal sessions. v2.1.274 added memory usage warnings and improved MCP server connectivity with configurable startup wait times. Recent updates also focused on prompt caching stability, improved artifact publishing, better error messages for various API failures, and enhanced terminal responsiveness. The releases include fixes for permission checking in Bash commands, improvements to fullscreen mode rendering, and numerous reliability improvements across MCP servers, plugins, and cloud sessions.