What Happened
OpenAI released its GPT-6 Astra model this week, calling it its most capable model yet. On benchmarks, it has new highs across computer use, coding, science, cybersecurity, and complex professional work.
Epoch’s Capabilities Index gives Astra a score of 169, up from 163 for Claude Fable 5.1 and roughly 137 for o1-mini two years ago. The frontier is plainly still moving.
The frontier has been advancing at roughly 14 ECI points per year. But in 2024 14 points represented about a 10.4% improvement. At Astra’s level, the same 14 points decline to about 8.4%.
In other words, absolute progress continues, while relative progress declines.
What It Means
- People experience relative improvement, not benchmarks.
Going from 50 to 70 can transform what a model can do. Going from 150 to 170 is much less likely to transform ordinary usage.
