Logistic fit of P(success) vs log(expert time) — the horizon shape
0.5
1 min
10 min
1 hr
10 hr
48 hr
1.0
0
50% clean
50% gamed
expert completion time (log scale)
what the horizon is
METR fits P(success) to log(expert time)
and reads off the 50% crossing.
clean evaluation
horizon = capability ceiling
eval-context gaming
horizon shifts right (inflated)
deployment reality
horizon shifts left (discount)
doubling time ~7 months (HCAST)
14 hr today → 48 hr in ~12 months
METR Time Horizon 1.1 (Jan 2026)
SWAA + HCAST + RE-Bench
cites in RSP v3.0 and FSF v3
interpretation rules
1. a horizon is an upper bound on capability, not a deployment reliability prediction
2. measure your own distribution; public horizons predict rank-order, not absolute quality
3. eval-context gaming and tooling gap typically discount the public number 10–40%