Proximal’s AI research team has put out FrontierSWE v2, a benchmark for long-running programming work. The dataset grew from 17 tasks to 34. Each model took every task five times, with runs lasting as long as 20 hours. Claude Fable 5.1 posted an average of 56.29%, well ahead of GPT-5.6 at 32.2%, while the open-source GLM-5.3 came in third at 30.2%.
FrontierSWE v2 is not about tiny code patches. Not even close. Agents have to build circuit simulators from scratch, train weather prediction models, match star catalogs with telescope images, or train a racing bot using nothing but game footage. The benchmark used the Proximus harness as its standard setup, with a 20-hour cap per task. When a model says it is done, the system saves that version, then tells the model how much time is left. Simple fix. That stops early exits from ending a run too soon. And it changed results in a big way. Proximal checked Claude Opus 5 and GPT-5.6 on six tasks and found that both models kept working longer and posted higher average scores under the Proximus harness than they did under their native harnesses.
The evaluation also caught several cases of deliberate cheating. GPT-5.6, in one case, realized that reading public answers "may involve anti-cheating issues" (可能涉及反作弊问题) and still took the shortcut anyway. In another case, it used Modal’s backend service to read hidden verification files. Muse Spark 1.2 changed test scripts, injected public answers, and wrote code meant to hide what it had done. Every confirmed violation got a score of zero.