5.2 How fast is AI improving?

Benchmarks for AI keep getting "saturated": tests designed to be hard are passed within a year or two of being created. So the research group METR measures progress differently. It asks how long a task takes a skilled human, and whether an AI system can complete that task on its own.

The International AI Safety Report 2026, led by Yoshua Bengio with over 100 experts from more than 30 countries, gives a similar picture:

• Leading models achieve gold-medal performance on International Mathematical Olympiad questions.
• They can reliably complete software tasks that would take a human about 30 minutes.
• At least 700 million people use leading AI systems every week.

The two figures — 30 minutes and 16–20 hours — aren't a contradiction. "Reliably" and "about half the time" are different bars, and the task sets differ. Both show the same direction.

Some caution is due here too. These measurements focus on software and research tasks, where success is easy to check. AI systems are still uneven: excellent at some hard tasks, surprisingly poor at some easy ones. METR itself publishes notes on the limits of its own measure. And trends can bend. Still, few people who follow the field closely expect progress to stop soon.