rotative
2026-09-14 · verified 2026-10-05 · re-checked weekly

the agi index

99 /100

time horizon

10m1h 8h workday gpt-4 4mo3 2hgpt-5 3.4hopus 4.6 12h17.4h †2023202420252026

50% task horizon, human-time, log scale · data: metr · † >16h

the four bars

arc-agi-2 · fluid puzzles · benchlm
95 vs 66 human avg · capped
gpqa diamond · phd science · intuitionlabs
94.1 vs 69.7 experts · capped
swe-bench verified · real github issues · vals
97, standardized scaffold
metr horizon · one 8h workday · metr
12.0h · 150% of bar · capped

gates 4/7

✓ novel reasoning ≥ human
95 vs 66
✓ computer use ≥ human
86.1 vs 72.4
✓ 8h autonomy
12.0h, 50% reliability
✓ engineering ≥ 90%
97
○ peer-reviewed discovery
none yet
○ economic substitution
no replicated evidence
○ interactive agentic reasoning ≥ human (arc-agi-3)
62.7% vs human 100% · semi-private, standard harness
method & caveats

each bar: best known model ÷ its human reference, capped at 100, equal weight, averaged. hle excluded (no human baseline). arc-agi-2 "at bar" is the public set; the real gate is 85% on the hidden set. swe-bench takes the standardized scaffold, not self-reports. the horizon point is opus 4.6, feb 2026, within metr's reliable range; newer points past 16h are marked hollow. benchmarks saturate and get replaced; replacements are announced, old series archived with the break marked. arc-agi-3 update: astra scores 62.7% on the semi-private set with the standard harness and 99.9% with the provider adapter — arc prize publishes both, and the harness changes the score.

← rotative