ResearchFriday, November 20, 2026· 2 min read

METR Update Shows AI Capabilities Still Advancing Fast

TL;DR

A new METR report updates its closely watched AI “time horizon” analysis, finding that some model capabilities continue to improve on an exponential path. The results also add nuance: AI appears stronger on coding-style tasks than on work requiring broader judgment, giving researchers a clearer map of where progress is happening.

Key Takeaways

  • 1METR’s updated time horizon plot suggests some AI capabilities are still improving rapidly.
  • 2The report highlights especially strong progress in coding-related tasks.
  • 3It also shows current models remain less reliable on tasks requiring judgment and discretion.
  • 4Better evaluation tools can help researchers, companies, and policymakers track real AI progress more accurately.

A new report from the AI evaluation nonprofit METR offers an important update to one of the field’s most debated charts: the AI “time horizon” plot. The updated analysis suggests that some AI capabilities are continuing to improve along an exponential trajectory, a sign that measurable progress is still moving quickly.

A clearer picture of AI progress

The findings are especially encouraging for understanding where today’s systems are becoming more useful. METR’s work indicates that models are performing particularly well on coding-style tasks, where success can often be tested and verified more directly.

At the same time, the report adds healthy nuance. Models appear less capable on tasks that require broader judgment, context, or discretion, reminding the field that progress is not uniform across every type of work.

  • Rapid gains: Some capabilities continue to follow a steep improvement curve.
  • Practical strength: Coding remains one of the clearest areas of model progress.
  • Better measurement: More precise evaluations help separate hype from real advances.

That makes this update a win for AI research: better benchmarks give everyone a more grounded way to understand what systems can do, where they are improving, and where more work is needed before deployment in higher-stakes settings.

Get the free AI Wins digest

Only positive AI stories, delivered daily or weekly. Free, no spam, unsubscribe anytime.