GPT-6 Luna: benchmark results
Intended for focused tasks repeated at volume. Supports structured outputs and tool calls, with reasoning that can be switched off for simpler requests. Provider: OpenAI. Released 2026-09-22. Access: API.
Aggregate Rating 1809 ± 5, rank #62 of 1640 rated models, from 895 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EpiSEM-Bench - Critical Unsafe Commit Rate | 21 | Score (%) | 100 |
| EpiSEM-Bench - Full Unsafe Commit Rate | 3 | Score (%) | 100 |
| EpiSEM-Bench - Recovery Unsafe Commit Rate | 3 | Score (%) | 100 |
| EpiSEM-Bench - Safety Boundary Margin | 84.8 | Score (%) | 100 |
| LoomBench (agents solo-accept; Chat Probabilities) - Attainment | 65.91 | Fraction of 44 instructed agents meeting their objective, in | 100 |
| LoomBench (agents solo-accept; Chat Probabilities) - Grow | 95.45 | Mean source-normalized Grow objective score (0-100), 11 agen | 100 |
| LoomBench (agents solo-accept; Chat Probabilities) - Hold | 77.73 | Mean source-normalized Hold objective score (0-100), 11 agen | 100 |
| LoomBench (agents solo-accept; Typed Decisions) - Give | 79.55 | Mean source-normalized Give objective score (0-100), 11 agen | 100 |
| LoomBench (agents solo-refuse; Chat Probabilities) - Attainment | 68.18 | Fraction of 44 instructed agents meeting their objective, in | 100 |
| LoomBench (agents solo-refuse; Chat Probabilities) - Grow | 90.91 | Mean source-normalized Grow objective score (0-100), 11 agen | 100 |
| LoomBench (agents solo-refuse; Chat Probabilities) - Hold | 78.64 | Mean source-normalized Hold objective score (0-100), 11 agen | 100 |
| LoomBench (agents solo-refuse; Typed Decisions) - Grow | 93.18 | Mean source-normalized Grow objective score (0-100), 11 agen | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-6-luna · How It Works · Data refreshed daily, snapshot 2026-11-07.