Astra and Fable 5.1 pull well ahead
GPT-6 Astra and Claude Fable 5.1 take the top four places in DuelLab's September 10 results. Their Medium entries already outscore every other officially ranked model variant. At XHigh, both move further ahead, although their strengths differ from game to game.
DuelLab asks AI models to write game-playing programs. Those programs are compiled and run against one another, and match results determine their ratings. The question is practical: can a model turn the rules of a game into code that plays it well?
This release covers eight games, 63 model families and 96,063 rated matches. The games range from Go and Chess to Hex, Game of the Amazons and Tumbleweed. Every number below comes from the same September 10 snapshot, so the comparisons share an evaluated field.
Medium is already enough to lead the rest of the ranked field
Astra Medium scores 79.4 and Fable 5.1 Medium scores 77.2. The strongest officially ranked entry from any other model is GPT-5.6 Sol XHigh at 75.4. The two new models therefore occupy first through fourth place when their Medium and XHigh entries are considered together.
The qualification matters: these are the official ranked entries. GLM 5.3 Max has a provisional score of 78.9, but has not qualified for that ranking in this snapshot. A comparison that silently included or excluded provisional rows would tell a different story.
Reasoning settings describe the generation of the player program. Once a match begins, that program makes the moves; the language model is not called on every turn. Medium and XHigh are the retained labels in this release, rather than equal amounts of computation across providers.
More reasoning helps the overall result, with some local reversals
Astra rises from 75.3 at Low to 79.4 at Medium and 91.0 at XHigh. Fable 5.1 follows the same overall direction: 73.8, 77.2 and 92.4. The increase from Medium to XHigh is 11.5 points for Astra and 15.2 for Fable.
That smooth overall progression hides uneven game results. Astra's Tumbleweed score drops from 81.5 at Medium to 68.5 at XHigh. Fable's Game of the Amazons score falls from 100.0 to 92.8. Both models improve overall even though their higher-setting programs do worse in particular games. The results do not identify the cause of those reversals; answering that would require inspecting the generated strategies and their matches.
The leading models have different strengths
At XHigh, Astra has the higher per-game score in five games and Fable in three. Fable still has the higher combined score. Its large advantages in International Draughts and Tumbleweed outweigh Astra's more numerous, generally smaller advantages elsewhere.
| Game | Astra | Fable 5.1 |
|---|---|---|
| Go | 89.1 | 77.0 |
| Chess | 98.9 | 100.0 |
| Shogi | 100.0 | 89.4 |
| Xiangqi | 98.3 | 88.6 |
| Hex | 100.0 | 91.2 |
| International Draughts | 74.2 | 100.0 |
| Game of the Amazons | 98.8 | 92.8 |
| Tumbleweed | 68.5 | 100.0 |
These are scores against the evaluated opponent pools, not a direct eight-game match between Astra and Fable. Scores are normalized within each game; 100 marks the top of that scale, not mathematically perfect play. The small combined gap of 1.4 points also deserves less emphasis than the different strengths visible underneath it.
The separate Overall Models view combines results across settings. There, Astra leads 81.9 to 81.1. This explains how the overall leader and the strongest individual reasoning variant can belong to different models: the views answer different questions.
A Hex player that won all 120 matches
The most concrete result is Astra's XHigh Hex player: 120 wins, zero draws and zero losses against 60 opposing programs. That is an undefeated record for one generated player in the published opponent pool. It does not mean Hex is solved, or that the player would beat a specialist Hex engine.
Hex has a compact objective: connect your two opposite edges with an unbroken chain of stones. Writing legal moves is only the beginning. A strong player must build connections while anticipating attempts to cut them. The match record gives us an observable outcome to investigate, rather than relying on how persuasive the model's explanation of its strategy sounds.
What the extra playing strength costs
Astra's estimated generation cost rises from $0.41 per game at Medium to $1.30 at XHigh. Fable 5.1 rises from $1.15 to $5.02. At XHigh, Fable's 1.4-point score advantage comes with about 3.9 times the estimated generation cost.
| Model | Setting | Score | Cost |
|---|---|---|---|
| GPT-6 Astra | Medium | 79.4 | $0.41 |
| GPT-6 Astra | XHigh | 91.0 | $1.30 |
| Claude Fable 5.1 | Medium | 77.2 | $1.15 |
| Claude Fable 5.1 | XHigh | 92.4 | $5.02 |
These are API generation estimates, including repair attempts, averaged with equal weight for each model, game and setting combination. They are not the cost of a match, a complete operating bill, or a measurement of equal compute. Once a player exists, it can take part in many matches without regenerating its code.
What this release adds
The September results show substantial separation between these new entries and the previous ranked field, including at Medium. They also show why a single ordering is incomplete: model choice, generation setting, game and cost all change the comparison.
Start with the Reasoning Variants standings, then inspect the per-game charts. The immutable release preserves the data used here as the live leaderboard evolves.