20-hour programming benchmark highlights performance gaps: Claude Fable 5.1 leads GPT-5.6 by more than 24 points, with GLM-5.3 ranking third.

Beating AI Express: AI research team Proximal has released FrontierSWE v2, a long-horizon programming benchmark. The number of tasks has expanded from 17 to 34, with each model running 5 trials per task, each trial taking up to 20 hours. Claude Fable 5.1 posted an average score of 56.29%, significantly outperforming GPT-5.6’s 32.2%. Open-source model GLM-5.3 ranked third with 30.2%. Tasks in FrontierSWE go far beyond basic code debugging: AI agents must build circuit simulators from scratch, train weather forecasting models, match star catalogs using telescope images, or train racing bots solely from game visuals. Version v2 has uniformly switched to the Proximus harness. Each task runs for up to 20 hours; when a model is ready to submit its work, the system saves its current state and notifies it of remaining time to prevent premature task termination—a change that significantly impacted scores. Proximal’s comparison across 6 tasks found that both Claude Opus 5 and GPT-5.6 ran longer with the Proximus harness, and posted higher average scores than with their original harnesses. The benchmark also caught multiple instances of intentional cheating: GPT-5.6 once recognized that accessing public answers “might involve anti-cheat issues” but still used the shortcut; on another occasion, it even exploited Modal’s backend services to read hidden validation files. Muse Spark 1.2 modified test scripts, inserted public answers, and wrote code to cover up its cheating traces. All runs confirmed to be violations were scored zero.