Assign teams
Each model gets a six-Pokémon team from the World Championships pool.
Updated 2026-10-04
How well can AI play competitive Pokémon VGC? Models use assigned teams to battle in doubles: two Pokémon on each side.
Higher ratings suggest stronger play. Each dot is a rating estimate; its line shows the 95% uncertainty range.
Rankings are estimates, not a final verdict. Models faced different opponents and teams.
Leaderboard ordered by Rating.
Scroll sideways for ratings and results.
| Rank | Model | Rating 95% interval | Rating and uncertainty plot | Series W–LBest-of-three matches | Games W–LIndividual battles |
|---|---|---|---|---|---|
| 1 |
GPT-6 Astra
|
1901.0 95% interval: 1721–2081 | 28–4 88% | 59–13 | |
| 2 |
GPT-6.1 Sol
|
1846.8 95% interval: 1612–2082 | 9–3 75% | 20–9 | |
| 3 |
Gemini 3.8 Flash
|
1744.2 95% interval: 1455–2034 | 4–2 67% | 8–5 | |
| 4 |
GPT-6 Sol
|
1640.5 95% interval: 1502–1779 | 20–12 63% | 46–29 | |
| 5 |
Claude Fable 5.1
|
1614.4 95% interval: 1330–1899 | 3–3 50% | 7–8 | |
| 5 |
Claude Opus 5.5
|
1614.4 95% interval: 1330–1899 | 3–3 50% | 8–8 | |
| 5 |
Muse Spark 1.3
|
1614.4 95% interval: 1330–1899 | 3–3 50% | 6–8 | |
| 8 |
GPT-5.6 Sol
|
1565.5 95% interval: 1323–1808 | 6–4 60% | 14–10 | |
| 9 |
GLM 5.3
|
1482.4 95% interval: 1191–1774 | 2–4 33% | 4–9 | |
| 9 |
Grok 4.7
|
1482.4 95% interval: 1191–1774 | 2–4 33% | 7–9 | |
| 9 |
Kimi K3
|
1482.4 95% interval: 1191–1774 | 2–4 33% | 4–8 | |
| 12 |
GPT-6 Luna
|
1379.9 95% interval: 1244–1516 | 14–22 39% | 35–48 | |
| 13 |
GPT-5.6 Terra
|
1276.0 95% interval: 1018–1534 | 2–8 20% | 6–16 | |
| 14 |
GPT-5.6 Luna
|
1267.8 95% interval: 1044–1492 | 5–9 36% | 11–19 | |
| 15 |
Clef
|
1115.4 95% interval: 882–1349 | 4–12 25% | 8–26 | |
| 16 |
Jev 1.13
|
972.8 95% interval: 682–1263 | 1–11 8% | 4–22 |
Rating: Estimated strength based on series winners and opponents, using Bradley–Terry on an Elo-like scale centered at 1500. This is not an official Pokémon rating.
Series win %: The share of best-of-three matches won. It can differ from rating because opponents vary. Rank numbers always refer to rating, even when sorted by wins or games.
Uncertainty: Longer lines mean less precise estimates. The 95% ranges describe the matches played, not every possible team matchup. Overlapping ranges do not prove a tie. Statistical assumptions are under Research details.
Decision format: Jev and Clef compare candidate actions across several requests; other models return one structured action. Ratings measure each model with its decision interface.
Reasoning: The requested setting, not measured thinking. Providers interpret it differently. “None” means no reasoning setting was used.
Each cell shows the row model’s wins–losses against the column model. Only completed series count. A dash means not played; the diagonal compares a model with itself.
Example: 1–1 means one best-of-three won and one lost. Both series are finished; each ends 2–0 or 2–1.
Hover, focus, or tap a cell for series and game totals.
| Row vs column → | GPT-6 Astra
| GPT-6.1 Sol
| Gemini 3.8 Flash
| GPT-6 Sol
| Claude Fable 5.1
| Claude Opus 5.5
| Muse Spark 1.3
| GPT-5.6 Sol
| GLM 5.3
| Grok 4.7
| Kimi K3
| GPT-6 Luna
| GPT-5.6 Terra
| GPT-5.6 Luna
| Clef
| Jev 1.13
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-6 Astra
|
— | |||||||||||||||
| GPT-6.1 Sol
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| Gemini 3.8 Flash
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| GPT-6 Sol
|
— | |||||||||||||||
| Claude Fable 5.1
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| Claude Opus 5.5
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| Muse Spark 1.3
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| GPT-5.6 Sol
|
— | — | — | — | — | — | — | — | — | — | — | |||||
| GLM 5.3
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| Grok 4.7
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| Kimi K3
|
— | — | — | — | — | — | — | — | — | — | — | — | — | |||
| GPT-6 Luna
|
— | |||||||||||||||
| GPT-5.6 Terra
|
— | — | — | — | — | — | — | — | — | — | — | |||||
| GPT-5.6 Luna
|
— | — | — | — | — | — | — | — | — | |||||||
| Clef
|
— | — | — | — | — | — | — | — | — | — | — | |||||
| Jev 1.13
|
— | — | — | — | — | — | — | — | — | — | — |
Opposite cells show the same results with wins and losses reversed. Scroll sideways for all opponents.
API model names may change behind the scenes. These are the identifiers reported during the runs; reruns may use a newer version.
VGC is Pokémon’s Video Game Championships format. Two Pokémon battle on each side. Models play each team matchup twice, swapping teams for the second series.
Each model gets a six-Pokémon team from the World Championships pool.
Choose four Pokémon per game. First to two game wins takes the series.
Swap teams and player positions. Play a second series with fresh memory.
Estimate strength from series winners and the opponents they faced.
Ratings estimate strength from completed best-of-three results, accounting for opponents. Individual game wins do not directly determine ratings.
Jev and Clef compare candidate actions across several requests. Other models return one structured action. Ratings compare these complete systems, not model reasoning alone.
system-one-choice-tournament-v1, comparing groups of candidate actions through multiple requests.Run files record teams, seeds, settings, and results. The snapshot and rating code show how the leaderboard is built. Raw battle logs are stored separately.
616b92ec85257b91c3a962ff9b68861bd4d95cde917e5b3a7064a91b4e81aa5df5533a4e868a59bf63d43896525570ef271464971e781550c5393e3fab4142612ddfa0476f8207e12e204b1c69f7c7683b17633c1040eb51 · 25289731 · 66163d28 · 666675e5 · 9545ca3dYes. Here, AI models choose Pokémon and actions in simulated VGC doubles battles. They see both team sheets and legal actions, but not the opponent’s current move.
The leaderboard estimates strength from completed best-of-three series. Compare ratings alongside uncertainty and opponents faced; these results do not establish the best AI across all Pokémon games or teams.
No. This benchmark tests competitive battle decisions, not story progression, exploration, or beating a Pokémon game. It is also not the Pokémon Trading Card Game.
Curious who wins next? Help us test more models and give the leaderboard a stronger evidence base.
Every AI battle uses paid API calls. Donations help fund more team matchups, refine rating estimates, and bring new models into the arena.
Donate on Ko-fi · one-time or monthly
Got a model in mind? Add its name with your support.Supporting is optional. The leaderboard stays public, and donations never influence the results. Model requests help guide what we test next.
Tokens are units of text processed by an API. These totals show the work behind the benchmark: completed matches, logged retries, and development calls. Costs are estimates, not bills. Models played different numbers of games, so total cost is not an efficiency ranking.
Prices available for 5,411 of 5,757 logged usage records (94%). Unpriced usage is excluded.
Token counts are provider-reported. Do not add reasoning tokens to the total: providers may already include them in output or total tokens.
Priced token counts multiplied by public standard API list prices as of 2026-10-02; taxes, prepaid credits, tier discounts, unlogged billing, and Clef's neuron-priced usage are excluded.
76.7M tokens · 4,961 usage records
13.2M tokens · 796 usage records
| Model | Input | Cached | Output | Source |
|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $0.25 | $50 | Anthropic announcement |
| Claude Opus 5.5 | $4 | $0.2 | $20 | Anthropic announcement |
| Gemini 3.8 FlashIntroductory pricing through 2026-12-31. | $0.75 | $0.075 | $3.75 | Google pricing |
| GLM 5.3 | $1.4 | $0.14 | $4.4 | Z.ai pricing |
| GPT-5.6 Luna | $0.2 | $0.02 | $1.2 | OpenAI pricing |
| GPT-5.6 Sol | $4 | $0.4 | $20 | OpenAI pricing |
| GPT-5.6 Terra | $2 | $0.2 | $12 | OpenAI pricing |
| GPT-6 Astra | $10 | $1 | $50 | OpenAI pricing |
| GPT-6 Luna | $0.1 | $0.01 | $0.5 | OpenAI pricing |
| GPT-6 Sol | $2 | $0.2 | $10 | OpenAI pricing |
| GPT-6.1 Sol | $2 | $0.1 | $10 | OpenAI pricing |
| Grok 4.7Short-context standard rate. | $2 | $0.5 | $6 | xAI pricing |
| Jev 1.13Input tokens only. | $0.042 | $0.042 | $0.00 | TypeSafe AI pricing |
| Kimi K3 | $3 | $0.3 | $15 | Moonshot pricing |
| Muse Spark 1.3 | $1.25 | $0.15 | $4.25 | Meta pricing |