GPT-5.6 Luna vs GPT-6 Luna
Compare with other modelsGPT-6 Luna takes a substantial step forward on unsupported claims. The share of conversations with no unsupported claims goes from 67% to 90%, while the cost falls from $0.0064 to $0.0040 per conversation. The overall quality score barely changes, from 85 to 87.
Grounding takes a step back, from 94 to 87, and median reply time goes from 3.9 to 4.8 seconds. If you can live with the latency increase, GPT-6 Luna offers great performance on supported claims at a very competitive price.
| Metric | OpenAI Sep 8, 2026 | OpenAI Sep 23, 2026 |
|---|---|---|
| Overview | ||
| Quality score | 85 #3 of 17 | 87 #1 of 17 |
| Median reply | 3.9s Best among selected models Served by OpenAI | 4.8s Served by OpenAI |
| Cost per conversation | $0.0064 | $0.0040 Best among selected models |
| Quality scores | ||
| Tool use | 85 Best among selected models | 83 |
| Task completion | 86 Best among selected models | 85 |
| Context retention | 86 Best among selected models | 85 |
| Grounding | 94 Best among selected models | 87 |
| Safety | 90 | 93 Best among selected models |
| Hallucinations % of conversations with no unsupported claims detected | 67% | 90% Best among selected models |
| Running stats | ||
| p95 reply | 9.5s Best among selected models | 14.0s |
| Consistency % of repeat runs ended the same way | 79% Best among selected models | 77% |
| Output per turn | 166 tokens, median | 267 tokens, median |
Reply times cover the model call only. Cost estimates use token usage at list prices.
This whole report is one Voxli workspace: simulated customers, assertion checks, and a frozen, versioned test set that reruns when new models ship.
Test set v3f-2026-09 · edition 2026-09-08