GPT-6.1 Sol debuts at #4 on GameBench 2
Five models join DuelLab's GameBench 2 in the October 6 results. The best of them, GPT-6.1 Sol, places fourth among 191 ranked model-and-setting entries at its XHigh reasoning setting. It scores 87.7, half a point behind Claude Fable 5.1 at XHigh. Its player programs cost an estimated $0.27 each to generate. For Fable 5.1 the figure is $5.02.
The top of the table
GPT-6 Astra at XHigh still leads with 93.1, and Claude Opus 5.5 at XHigh follows with 90.7. The next three entries, within five points of each other, include two newcomers at XHigh, GPT-6.1 Sol and Claude Sonnet 5.5.
| Rank | Model and setting | Score | Est. cost per player |
|---|---|---|---|
| 1 | GPT-6 Astra · XHigh | 93.1 | $1.30 |
| 2 | Claude Opus 5.5 · XHigh | 90.7 | $1.99 |
| 3 | Claude Fable 5.1 · XHigh | 88.3 | $5.02 |
| 4 | GPT-6.1 Sol · XHigh (new) | 87.7 | $0.27 |
| 5 | Claude Sonnet 5.5 · XHigh (new) | 83.6 | $0.55 |
In the Overall view, which averages each model's three tested settings, GPT-6 Astra leads with 79.4, followed by Claude Opus 5.5 (77.6), Claude Fable 5.1 (76.1) and GPT-6.1 Sol (73.6). The next model, Claude Opus 5, is far behind at 59.1.
Five new models
All five are ranked at each of their three tested settings. The table orders them by their best single setting. The Overall rank averages all three.
| Model | Best setting | Score | Overall rank |
|---|---|---|---|
| GPT-6.1 Sol | XHigh | 87.7 | 4 |
| Claude Sonnet 5.5 | XHigh | 83.6 | 11 |
| Gemini 3.1 Pro Preview | High | 49.3 | 40 |
| Ember-1 | Max | 46.1 | 30 |
| Space Bunny Alpha | XHigh | 34.4 | 52 |
No model in the benchmark changes more with its reasoning setting than Claude Sonnet 5.5. At XHigh it places fifth with 83.6. At Medium it scores 36.4 (70th), and at Low 28.1 (125th). Its Chess player alone goes from 7.4 at Medium to 96.6 at XHigh. Each setting has one player per game, so some of a swing that size can be the luck of a single generation.
Three players without a loss
Three generated programs did not lose a single rated match against the published field:
- GPT-6 Astra's XHigh Hex player: 382 wins, no draws and no losses against 149 different opponents.
- Claude Opus 5.5's XHigh International Draughts player: 251 wins, one draw and no losses against 126 opponents.
- Claude Sonnet 5.5's XHigh Tumbleweed player: 292 wins, two draws and no losses against 147 opponents.
These records describe play against other AI-written programs in this release. They do not mean a game is solved or that a player would beat a specialist engine.
What the strongest players cost
Generation cost is the estimated list price of producing one player program for one game. Near the top it varies by more than a factor of ten.
| Model and setting | Score | Est. cost per player |
|---|---|---|
| GPT-6.1 Sol · Medium | 73.6 | $0.08 |
| GPT-6.1 Sol · XHigh | 87.7 | $0.27 |
| Claude Opus 5.5 · Medium | 82.9 | $0.33 |
| GPT-6 Astra · Medium | 76.6 | $0.41 |
| Claude Sonnet 5.5 · XHigh | 83.6 | $0.55 |
| Claude Fable 5.1 · Medium | 71.5 | $1.15 |
| GPT-6 Astra · XHigh | 93.1 | $1.30 |
| Claude Opus 5.5 · XHigh | 90.7 | $1.99 |
| Claude Fable 5.1 · XHigh | 88.3 | $5.02 |
Claude Opus 5.5 at Medium is the most even of the strong entries. Its weakest game scores 77.3 and its best 90.4. Costs are list-price estimates, not invoices, and the data marks Claude Fable 5.1's figures as incomplete. Scores are relative to the current field, not absolute quantities, so they should not be divided by cost.
How to read this release
- The results include 347,619 rated matches between 1,740 player programs, written by 78 models for eight games.
- Ratings now come from a Glicko-style fit. Scores and ranks from earlier releases used sequential Elo and should not be compared one-to-one with these. The site shows no rank-change arrows across the change.
- The results pool matches played under two versions of the program runtime, so this release shows no "Interval validated" badges.