Bijan Bowen tests GPT-6.1 Sol through a browser desktop, games and interactive 3D scenes. He varies reasoning effort and speed settings, then reruns a pool-party task without fast mode because the initial comparison was not equivalentEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
The browser desktop and skating game are functional but less elaborate than his recent Sonnet 5.5 results. A follow-up improves the skating sceneIterative refinement improves an AI result through repeated cycles of building, inspection, feedback, and revision. and adds a successful replay camera, while the subway shooter follows the revised instructions and produces a more convincing result.
A detailed recreation of a television apartment, a RuneScape-style interface and an interactive car model show stronger object and interface detail. Important inaccuracies remain in spatial layout, game mechanics and the car's shapeCoding output quality measures whether AI-generated software is correct, maintainable, complete and suitable for its intended use.; these demonstrations are qualitative tests rather than a comprehensive benchmarkA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions..
The robot-arm experiment navigates more cautiously but does not complete its taskAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work.. A legacy-laptop experiment also fails and partly switches to a different model, so it cannot establish Sol's capability there. Bijan Bowen reports modest subscription usage, while explicitly noting uncertainty about whether plan allowances had changed.
Watch on YouTube




