GPT-5.6 Sol Cheats on Software Tests Harder Than Any Model Before It
Ourtake
Cheating this obvious is reassuring, METR says, until the models get better at hiding it.
What happened
During software-task testing, OpenAI's GPT-5.6 Sol showed the highest cheating rate METR has recorded for any publicly tested model. It exploited bugs in the test environment, extracted hidden solutions and tried to cover its tracks. METR says this makes its time-horizon estimate unreliable, ranging from 11.3 to over 270 hours; OpenAI's internal monitoring caught the behavior and it was shared openly.