Pareto 26.10 Preview debuts in the top ten on GameBench 2
Pareto 26.10 Preview, a multi-model AI system from Circuit & Chisel offered through its Unbiased AI platform, joins DuelLab's GameBench 2 in the October 13 results. It places seventh among 217 ranked model-and-setting entries, with a score of 78.2. Its player programs cost an estimated $0.03 each to generate, less than any other entry in the top thirty.
Where it lands
Pareto 26.10 Preview is offered with a single setting: reasoning controls are not available, so it was tested once, at the provider's default. That single setting sits between the Medium settings of Claude Opus 5.5 and GPT-6 Astra.
| Rank | Model and setting | Score | Est. cost per player |
|---|---|---|---|
| 6 | Claude Opus 5.5 · Medium | 83.4 | $0.33 |
| 7 | Pareto 26.10 Preview · Provider default (new) | 78.2 | $0.03 |
| 8 | GPT-6 Astra · Medium | 75.7 | $0.41 |
| 9 | GPT-6.1 Sol · Medium | 73.1 | $0.08 |
| 10 | Claude Opus 5 · XHigh | 71.9 | $1.38 |
The top five are the XHigh settings of Claude Opus 5.5 (91.8), GPT-6 Astra (90.7), GPT-6.1 Sol (87.3), Claude Fable 5.1 (86.2) and Claude Sonnet 5.5 (83.5). Costs are estimates, and the data marks Claude Fable 5.1's figures as incomplete.
Top-ten play at a fraction of the cost
Pareto 26.10 Preview's player programs cost an estimated $0.031 each to generate, or $0.25 for all eight games. No entry scores higher for the same money or less. The six entries ranked above it cost from nearly nine times as much per program (GPT-6.1 Sol at XHigh, $0.27) to 164 times as much (Claude Fable 5.1 at XHigh, $5.02). The other nine entries in the top ten cost a median of $0.55 per program, about 18 times Pareto's.
Seen from the other side, 75 other ranked entries cost the same or less per program. The best of them scores 50.8, about 27 points lower. The two entries just below it, GPT-6 Astra at Medium (75.7) and GPT-6.1 Sol at Medium (73.1), cost $0.41 and $0.084 per program: about 13 times and nearly three times as much.
Pareto 26.10 Preview's figure is what the provider charged. Most other entries' costs are estimates from token counts and list prices, so small differences should not be over-read. These gaps are large.
No Overall rank
The Overall view averages each model's three reasoning settings, including the weaker, low-reasoning ones. Pareto 26.10 Preview has a single setting, so an Overall average for it would not be comparable. It is listed below the Overall leaderboard without a rank, and the like-for-like comparison is the seventh place above.
Game by game
It won 3,000 of its 3,420 rated matches, drew 127 and lost 293, against 228 different opponents. It placed in the top ten in six of the eight games:
- Game of the Amazons: 92.7, seventh of 213 entries (453 wins, 17 losses).
- Chess: 77.3, eighth of 197.
- Shogi: 80.4, eighth of 191.
- Hex: 86.4, eighth of 212.
- Xiangqi: 69.0, tenth of 203.
- Tumbleweed: 83.9, fourth of 216 (221 wins, 11 losses). Its matches there are still being played, so this can move.
It was weaker at International Draughts (15th of 206) and Go (39th of 209).
How to read these results
- One player program per game stands behind the score, so a single strong or weak generation moves it noticeably.
- Costs are for generating one player program for one game: the provider's charge for Pareto 26.10 Preview, and list-price estimates from token counts for most other entries. They are not a measure of how well the program plays.
- Scores are relative to the current field of AI-written programs, and this is a preview model whose behavior may change before a final release.
- Its Tumbleweed program was first rejected by a fault in DuelLab's admission speed check. After the fix it was re-checked unchanged and admitted, so it joined the round robin later than the other programs.