Debating ExploitBench: would 96.8% from a flash model break the nonlinear-intelligence thesis?
teortaxesTex · x · 2026-10-31
teortaxesTex dissects Lisan al Gaib's nonlinear-returns-to-intelligence thesis using ExploitBench numbers: GPT-5.5 only improved 45%→72% going from 300 to 3000 turns, while at max=300 turns GPT-5.6 Sol scores 78.5%, Fable 5.1 83%, Opus 5.5 91%, Kimi K3.1 Pro 74.7%. A flash model at 96.8% (screenshots show GPT-6.1 Terra at 97.1%, with acknowledged contamination risk) would be extraordinary and challenge the thesis.
More from Models
- ThursdAI crew debates: with Opus 7 coming, who still cares about better benchmarks? — thursdai_pod · 2026-10-31
- 14 models tested on 960 analyst tasks: AI scores 97/100 as analyst, picks winners no better than random — HeyAmit_ · 2026-10-31
- Anthropic adds invisible text watermark to Claude outputs starting Nov 3 — petesena · 2026-10-31
- GPT-6 Luna vs GPT-6.1 Terra vs GPT-6.1 Sol vs GPT-6 Astra in Pixel Art — Angaisb_ · 2026-10-31
- Perplexity engineer invites public tasks to prove or fix pplx-decider's SOTA claims — antoine_chaffin · 2026-10-31
- Arena weekly: Gemini 4 Carbon tops Text Arena, GPT-6.1 Terra enters Agent Arena at #8 — arena · 2026-10-31