Ben Todd: AI Benchmarks Keep Beating Forecasts Across the Board

On September 10, 80,000 Hours founder Ben Todd published a thread systematically reviewing the AI progress predictions he made in August 2025. His conclusion: nearly every metric was far surpassed by actual progress, and people have generally underestimated the pace of AI development.

Confirmed

FrontierMath: Todd predicted 32% by end of 2026 and 97% by 2030, which was already on the optimistic side; in reality it hit 40% by the end of 2025 and now stands at 94%.
LiveCodeBench Pro: Already reached 53% in 2025, far exceeding Todd's own forecast of 23%.
METR 80% task-length benchmark: Most forecasters expected the 3-hour milestone only by end of 2026; it was actually achieved in April 2025, roughly 20 months early. Todd's prediction of 6 hours by year-end looks on track.
AI revenue: Superforecasters predicted $90 billion by end of 2026; in reality it passed $100 billion in August 2025, making it the fastest-moving metric of all. Todd considers his own $225 billion forecast the only one that might turn out too optimistic.
Share of AI-assisted tasks: Most forecasters expected about 4% of tasks to be AI-assisted by end of 2025; the actual figure was 5.7%. His forecast of 12% for 2027 may still pan out.

Why it matters

Todd's core point: even optimistically inclined forecasters in the field (himself included) systematically underestimated AI progress on both benchmarks and commercialization, suggesting future forecasts should be more aggressive.
He also notes a macro-level exception: GDP growth has yet to take off due to AI, in line with other forecasters' views—micro-level capability and business metrics exceeding expectations have not yet translated into significant macroeconomic change.

2026-09-10 ~ 2026-09-10 · 5 related posts

Primary sources

Ben Todd: AI hit 94% on FrontierMath, far outpacing even his bullish 32%-by-2026 forecast — ben_j_todd · 2026-09-10 11:01 (UTC+08:00)
Ben Todd: AI-assisted tasks hit 5.7% vs 4% expected, but GDP growth still flat — ben_j_todd · 2026-09-10 11:01 (UTC+08:00)
METR 3-hour task horizon hit 20 months early; LiveCodeBench Pro already at 53% — ben_j_todd · 2026-09-10 11:01 (UTC+08:00)

[source] Ben Todd: AI hit 94% on FrontierMath, far outpacing even his bullish 32%-by-2026 forecast — ben_j_todd · 2026-09-10
Ben Todd's forecast recap: FrontierMath at 94%, LiveCodeBench Pro at 53% — all beaten — ben_j_todd · 2026-09-10
[source] METR 3-hour task horizon hit 20 months early; LiveCodeBench Pro already at 53% — ben_j_todd · 2026-09-10
AI revenue already topped $100B, beating superforecasters' $90B end-2026 call — ben_j_todd · 2026-09-10
[source] Ben Todd: AI-assisted tasks hit 5.7% vs 4% expected, but GDP growth still flat — ben_j_todd · 2026-09-10