First Proof Batch 3 — verification council 9/10, solo model 7/10, at about 52x the cost
| Date | 2026-10-16 |
|---|---|
| Category | Benchmarks |
| Judgement | Neutral |
| Gauge weight | 1 / 3 |
| Source | First Proof Batch 3 (announced 2026-10-14, reported by TechTimes 2026-10-16), recorded in AGI watch, 2026-10-16 |
Summary
On mathematics proof tasks, the ProofCouncil structure scored 9/10 while a solo GPT-6 Astra scored 7/10. The open-weight Humanize scored 0/10, assessed as roughly ten months behind, and the council structure cost $1,888 versus $36 — about 52 times more.
A/B Impact Analysis
Heterogeneous verification demonstrably working fits Scenario A's direction, but the result was achieved on human-given problems at 52 times the cost, which simultaneously shows the distance to autonomous research. The debate judged it as satisfying only half of Crux B, so it is logged as neutral. Whether solo performance catches up to the council as the cost curve falls is the follow-up watchpoint; until then it decides for neither side.