Can LLMs play Magic, and what does skill cost?
Frontier models play full games of Magic: The Gathering on the XMage rules engine, against a fixed CPU opponent and each other. Every game is rated on one scale and priced by the tokens it used.
Skill per dollar
Hover or tap a point for detailsHigher is stronger, further left is cheaper. Elo is measured against the XMage CPU (0); the right axis translates it into the chance of beating the CPU in one game. Shaded capsules are 90% intervals; lines join one model’s reasoning-effort settings; the gold band traces the best rating available at each price.
What the games show
88% chance it is the strongest setting tested, from 10 games. Its interval is wide, so read it as the likely leader rather than a settled one.
84% likely stronger, and it won 3 of their 4 games head to head, at $0.51 a game against Sonnet's $1.77.
Haiku 5.5 at low and medium effort play level (−66, −73); at high it reaches +147, level with Luna @xhigh, for $0.22 a game.
GPT-6 Luna climbs from −137 with no reasoning to +134 at xhigh, every step under $0.07 a game: the cheapest way onto this chart.
Watch the games
All 70 games →Leaderboard
How it’s measured →| # | Model | Elo | 90% interval (CPU at the dotted line) | Beats CPU | P(#1) | Games | $ / game |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6.1 Sol@high | +709 | 98% | 88% | 10 | $0.85 | |
| 2 | GPT-6.1 Sol@medium | +265 | 82% | 2% | 16 | $0.51 | |
| 3 | GPT-6 Astra@medium | +229 | 79% | 2% | 8 | $2.64 | |
| 4 | Claude Opus 5.5@medium | +226 | 79% | 5% | 4 | $3.91 | |
| 5 | Claude Haiku 5.5@high | +147 | 70% | <1% | 34 | $0.22 | |
| 6 | GPT-6 Luna@xhigh | +134 | 68% | <1% | 75 | $0.060 | |
| 7 | Claude Sonnet 5.5@medium | +129 | 68% | <1% | 12 | $1.77 | |
| – | Claude Fable 5.1@medium | unrated | 63% | 3% | 2 | $10* | |
| 8 | GPT-6 Luna@high | +57 | 58% | 0% | 105 | $0.043 | |
| 9 | GPT-6 Luna@medium | +35 | 55% | 0% | 87 | $0.035 | |
| 10 | GPT-6 Luna@low | −61 | 41% | 0% | 102 | $0.023 | |
| 11 | Claude Haiku 5.5@low | −66 | 41% | 0% | 35 | $0.12 | |
| 12 | Claude Haiku 5.5@medium | −73 | 40% | 0% | 33 | $0.17 | |
| 13 | GPT-6 Luna@none | −137 | 31% | 0% | 49 | $0.036 |
Ratings fit 412 games (2026-10-07 to 2026-10-11): games against the CPU and model-vs-model battles on one Bradley–Terry scale, with deck strength, seat and the luck of each seeded deal separated out. Costs are recorded tokens at OpenRouter list prices. * Measured before a prompt-caching fix that cut Claude costs by up to 2.6×; likely overstated.