You’re reading the October 2026 report.
See November 2026, the latest → October 2026 baseline
Does the standard hold on each new generation?
A generation is compared only with its predecessor in the same run: same requests, same judges, same standard.
Each change gets one of three verdicts, and it counts only when its interval clears zero and it's at least
5 points with guidance or 10 points without.
Same-run pairs1
First looks8
Judge anchor checkNot set up yetStarts with the November run
Same model, month to monthNot reported yetWaits for the noise-floor measurement
Generation pairs
Predecessor Successor Too small to count 95% interval on the change
OpenAI · GPT Luna line · lineage proposed, confirm before publishing Human-centered Without guidance No detectable change
Need revision, 0 to 60% GPT-5.6 Luna 7 of 68 (10.3%), GPT-6 Luna 4 of 68 (5.9%) Change in points, −40 to +40 −4.4 points, interval −13.2 to +2.9. Counts only beyond ±10. Human-centered Human-centered guidance No detectable change
Need revision, 0 to 60% GPT-5.6 Luna 1 of 68 (1.5%), GPT-6 Luna 0 of 68 (0.0%) Change in points, −40 to +40 −1.5 points, interval −4.4 to 0.0. Counts only beyond ±5. Trauma-informed Without guidance No detectable change
Need revision, 0 to 60% GPT-5.6 Luna 5 of 24 (20.8%), GPT-6 Luna 2 of 24 (8.3%) Change in points, −40 to +40 −12.5 points, interval −29.2 to 0.0. Counts only beyond ±10. Trauma-informed Trauma-informed guidance No detectable change
Need revision, 0 to 60% GPT-5.6 Luna 3 of 24 (12.5%), GPT-6 Luna 2 of 24 (8.3%) Change in points, −40 to +40 −4.2 points, interval −12.5 to 0.0. Counts only beyond ±5. What this run can see
With 68 drafts per model per condition, a change has to be about 20 points to show reliably when about 30% of drafts need revision,
and about 7 points when about 2% do. So the Benchmark can catch a generation that falls apart under guidance. It can't separate two models
that both sit near zero, and it says so instead of calling a winner.
| Model | Lab | Snapshot | Days before the run | Generation verdict |
| Claude Opus 5.5 | Anthropic | Sep 21, 2026 | 29 | No tested predecessor; placement only |
| Claude Sonnet 5.5 | Anthropic | Sep 28, 2026 | 22 | No tested predecessor; placement only |
| DeepSeek V4.1 Flash | DeepSeek | Oct 17, 2026 | 3 | No tested predecessor; placement only |
| Gemini 4 Argon | Google | Oct 13, 2026 | 7 | No tested predecessor; placement only |
| GPT-6 Luna | OpenAI | Sep 22, 2026 | 28 | Paired with GPT-5.6 Luna, same run |
| GPT-6.1 Sol | OpenAI | Sep 29, 2026 | 21 | No tested predecessor; placement only |
| Muse Wonder | Meta | Oct 12, 2026 | 8 | No tested predecessor; placement only |
| Nemotron 3.5 Super VL | NVIDIA | Oct 15, 2026 | 5 | No tested predecessor; placement only |