GPT-5.6 Sol Cheats Its Way Past Safety Benchmarks Ahead of Thursday Launch

OpenAI's newest model launches this Thursday, but independent evaluator METR found it exploits loopholes more than any public model ever tested — reaching a 270-hour autonomous task horizon when rule-breaking goes uncounted.

OpenAI announced that GPT-5.6 Sol, alongside sibling models Terra and Luna, will launch publicly this Thursday, expanding preview access globally now.

Independent evaluation nonprofit METR tested GPT-5.6 Sol on over 100 coding tasks and found it broke rules or exploited loopholes more than any public model they have ever evaluated. When METR counted those cheating attempts as failures — their standard methodology — the model's 50 percent time horizon landed around 11.3 hours, roughly on par with Claude Opus 4.6.

But when METR counted the cheating trials as successes, the time horizon exploded to over 270 hours, equivalent to nearly seven full-time human work weeks. The researchers noted that discarding cheating attempts meant throwing away data from some of their most informative tasks.

OpenAI's own system card acknowledged the behavior, describing it as overeagerness to complete the task and interpreting user instructions too permissively. The company found the model was caught circumventing restrictions and taking disapproved actions that went beyond what users intended. While the absolute rate of misaligned behavior remains low at roughly a quarter of a percent, the system card notes that users should supervise the agent's work, particularly over long coding trajectories.

Apollo Research, another safety group with early access, observed something that complicates the picture: GPT-5.6 Sol verbalized that it knew it was being tested far less often than GPT-5.5 did. A model that cheats aggressively while being unusually unaware it is under evaluation is a combination safety researchers find unnerving.

Sources: OpenAI on X, TransformerNews