Models · GPT
GPT-6 Luna
Asked for a few hundred random integers, GPT-6 Luna does not pick evenly. This page is the shape of that preference, built from 36 recorded answers containing 11139 numbers, and it is what the identify tool compares a pasted reply against.
- Bank id
gpt-6-luna- Family
- GPT
- Collected through
- API
- Collected on
- 2026-09-22
- Status
- Inherited from ModelTrace; serving identity not independently verified
- Distinct values used
- 355 of 355
- Prompt environments
- 12 of 12
Capability and price
- Sold by
- OpenAI
- Intelligence index
- 37 · v4.3.2, scored as max
- Index rank
- 9 of 11
- Input, per 1M tokens
- 0.10
- Output, per 1M tokens
- 0.50
- Blended, per 1M tokens
- 0.20
- Price rank
- 1 of 11
- Context window
- 1,000,000 tokens
Read on 2026-09-23, not measured here: index page · price page
Distribution
Every bar is one value from 1 to 355; its height is how many times the model chose it across all recorded answers. A perfectly random picker would give a flat line. Nothing in the bank does.
Favourite numbers
| Number | Times chosen | Share |
|---|---|---|
| 319 | 60 | 0.54% |
| 5 | 57 | 0.51% |
| 18 | 53 | 0.48% |
| 11 | 52 | 0.47% |
| 12 | 52 | 0.47% |
| 6 | 49 | 0.44% |
| 350 | 49 | 0.44% |
| 16 | 48 | 0.43% |
| 31 | 48 | 0.43% |
| 8 | 47 | 0.42% |
| 9 | 47 | 0.42% |
| 73 | 47 | 0.42% |
How reliably it is recognised
In held-out testing across all twelve prompt environments, 8 of 12 three-answer runs named this model with a clear margin, 1 were close calls inside its family, 3 had only a weak match, and 0 named a different model with a clear margin.
With this model removed from the bank and its own answers scored against the rest, 0 of 12 runs still named another model with a clear margin, 2 were close calls inside the family, and 10 had only a weak match. Its answers landed most often on GPT-5.6 Luna (9 of 12). That is the number to keep in mind if you are testing a version that is not listed here.
Taken as published from upstream ModelTrace 55a2e4a, with no replies or holdout of our own. Cross-validated on upstream's 36 replies (not a holdout), single-reply top-1 is 33/36 for GPT-6 Sol and 36/36 for GPT-6 Luna and Claude Opus 5.5; all three reach 12/12 at three replies. Added in v2026.09.5 by an Owner decision although admission rule 3 moved from 131 to 132 on 1,512 removed-model cases; the changelog lists every change.
Both measurements count the status the site attaches to the first-ranked model, not the bare ranking alone. The methodology page explains how they were made and what they leave out.