METR: AI models' autonomous work horizons keep doubling roughly every 3.6 months, but research judgment lags
METR published its Q4 capability update (Time Horizon 2), evaluating 15 frontier model checkpoints. It found that models' 50%-success task horizons on verifiable coding tasks are doubling about every 3.6 months, while performance on open-ended research judgment remains much weaker. METR says current evidence cannot rule out drastic AI R&D acceleration within a couple of years.
Why it matters: If the doubling trend holds, the length of work agents can finish unsupervised keeps compounding — a planning input for every team adopting them.
Who should care: AI developers, researchers, and policy watchers
OpenAI says new safeguards cut reasoning-extraction attacks on GPT-6.2 Astra from 54% to 0.2%
OpenAI published research on adversarial distillation — rivals extracting a model's hidden reasoning to train their own. In OpenAI's internal red-team tests, attacks pulling protected reasoning succeeded 54.0% of the time against GPT-6 Astra but 0.2% against GPT-6.2 Astra, after OpenAI tied protected reasoning to the conversation history that produced it. The figures are OpenAI's own evaluations.
Why it matters: Protecting hidden reasoning is becoming a front line in the model-copying fight, and it may shape how much capability leaks between labs.
Who should care: AI developers, security teams, and model providers