GPT-6 Luna, GPT-6.1 Terra and GPT-6.1 Sol each answered 25 of 25 checkable tasks correctly (100%), so these questions no longer tell them apart.
Seonara model bench, 2026-11-04: 25 tasks x 3 run(s) at temperature 0, task set 2026-10-29.1, every answer scored by a program.
| Model | Correct | Change | dates-units | extraction | instructions | knowledge | reasoning | Median seconds | Cost per run | Consistent |
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-6 Luna | 25 of 25 (100%) | new | 100% | 100% | 100% | 100% | 100% | 1.28 | $0.0005 | 84% |
| GPT-6.1 Terra | 25 of 25 (100%) | new | 100% | 100% | 100% | 100% | 100% | 1.32 | $0.0036 | 92% |
| GPT-6.1 Sol | 25 of 25 (100%) | new | 100% | 100% | 100% | 100% | 100% | 1.80 | $0.0061 | 96% |
| GPT-4.1 | 24 of 25 (96%) | 0 | 75% | 100% | 100% | 100% | 100% | 0.86 | $0.0028 | 92% |
| GPT-4o mini | 23 of 25 (92%) | 0 | 75% | 100% | 80% | 100% | 100% | 0.81 | $0.0002 | 100% |
| GPT-4.1 mini | 22 of 25 (88%) | 0 | 75% | 100% | 60% | 100% | 100% | 0.79 | $0.0006 | 96% |
Since 2026-10-29
No model's score moved.
Consistent: the share of questions a model answered the same way in every one of its 3 runs; a model that changes its answer is not one to rely on for that question.
On the harder questions
The same models on a second set of 7 tasks built to trip them (task set 2026-10-29.leap1, 3 run(s) each, run 2026-11-04). A trap task has a tempting wrong answer; a control task has none.
| Model | Correct | control | trap | Median seconds | Cost per 100 tasks |
|---|---|---|---|---|---|
| GPT-6.1 Terra | 100% | 100% | 100% | 1.58 | $0.0400 |
| GPT-6.1 Sol | 100% | 100% | 100% | 2.23 | $0.0648 |
| GPT-6 Luna | 86% | 83% | 87% | 2.12 | $0.0076 |
| GPT-4.1 | 67% | 100% | 53% | 0.93 | $0.0566 |
With only 7 tasks one miss is 14 points before repeats; read gaps of a few points as ties.
What this cannot tell you
25 questions is a small sample: a one-question difference is 4 points, so models within a few points of each other are tied. Seconds are wall time from one PC, network included. Costs use the listed prices per million tokens on the day. The questions, the scoring code and every answer are published, so the run can be repeated.
Raw results: results.csv and methodology.md of the 2026-11-04 run, published with this scoreboard.