METR’s latest evals show GPT-5.2 (High) hitting a "50% time-horizon" of 6.6 hours. For people who don’t know, their methodology measures autonomous capability by timing how long a human expert takes to finish a task that the AI can now complete successfully half the time. I spoke of "Light Takeoff" with OpenAI and this is further proof. This trajectory is the definition of Light Takeoff - a smooth but real exponential acceleration. The capability doubling time has shrunk from 212 days to just 128 days. This 6.6-hour record was set using "High" reasoning effort, not the even more powerful "XHigh" setting. With GPT-5.2 already beating Claude 4.5 Opus (5.3 hours) by 25%, the "XHigh" version likely pushes the horizon past 10+ hours. This proves OpenAI is successfully scaling reasoning compute. And this isn’t even 5.3!