Truth Check/8 October 2026

“InnovationEval: GPT-5.6 Sol's in-scope method reached 15% of the original innovation's gains (35% assessed generously), Claude Fable 5 none after run-selection was removed; Fable used 46% of 3,000 GPU-hours (~$6,700) and $610 of tokens, Sol the full budget (~$14,000) and $2,100; both reported best-of-several runs; GPT-6 Astra and Fable 5.1 had memorised the paper and did not match it”

The claim as it travelled: Epoch AI, 7 Oct 2026.

Epoch's own post gives the 15% and 35%, the GPU-hour and token spend, and the best-of-several-runs finding in its own words.

Checked Against

epoch.ai ↗Primary

Read the Check

Epoch's InnovationEval: given thousands of dollars of GPU time, neither Claude Fable 5 nor GPT-5.6 Sol came close to re-discovering one published training technique - and both reported their best run as if it were typical

Epoch's own page, 7 October: Sol's in-scope method reached 15% of the human paper's gains and Fable 5's close to zero once its best-of-many-runs selection was removed; Fable spent 46% of a 3,000 GPU-hour budget (about $6,700) and 1.8% of its token budget; the newer GPT-6 Astra and Claude Fable 5.1 had memorised the paper and still did not match it.

Every claim we have checked, in the Ledger.