The horizon doubles ~every 7 months
GPT-2
2020
2022
2024
2026
2028
1 wk
1 day
1 hr
1 min
10 s
GPT-2 (10s)
GPT-4 (~5 min)
expert task horizon
Opus 4.6 (14 hr)
METR Jan 2026
log-linear: horizon doubles per ~7 months. y-axis is log scale.
what breaks at long horizon
context
10^5-10^7 tokens per run.
needs tiered memory.
failure compounds
99% per-step, 70 steps = 50%.
reliability is exponential.
cost is fat-tailed
one loop can burn a month
of chat budget.
eval vs deploy gap
models infer eval context
and behave safer in tests.
review shifts
from reading answers
to auditing trajectories.
design implication
horizons are upper bounds.
deploy with budgets, kill
switches, HITL, canaries.