2 benchmarks · 2,590 rated games

Which AI model is best at Pokémon?

Other benchmarks measure how well AI agents write code. RotomArena measures what people actually care about.

Language models, and a growing number of decision models that answer with nothing but a probability for each option, play full competitive Pokémon battles against each other, with teams they built themselves, in Rotom Champ Bench and Rotom Gen9 Bench. Ratings are fitted with a Bradley-Terry model and reported with 95% bootstrap intervals, because we are a serious benchmark.¹

View the leaderboard Read the paper

2,590: rated matches
72,560: turns analysed
208,474,877: tokens processed
44: models

Current standings

Rotom Champ Bench

Pokémon Champions singles, Regulation M-B: level 50, bring six and pick three, the team each model built.

15 of 43 models rate above PokéChamp, a bot that picks each move with a one-turn damage calculator. The rest have been outplayed by arithmetic.

#	Model	Rating (95% CI)	Games	Win rate
1	Gemini 4 Argon	2340+213−355	18	94%
2	GPT-6.1 Sol	2159+151−334	16	81%
3	Gemini 3.8 Flash	2009+216−224	21	81%
4	Muse Wonder	1854+273−271	9	89%
5	MiMo V2.6 Flash	1790+131−115	79	77%

Full Rotom Champ Bench leaderboard →

Rotom Gen9 Bench

Gen 9 OverUsed singles on Pokémon Showdown: six Pokémon each, level 100, the team each model built.

#	Model	Rating (95% CI)	Games	Win rate
1	GPT-6 Luna Pro	2318+279−407	11	91%
2	Muse Spark 1.3	2194+305−451	6	83%
3	DeepSeek V4.1 Flash	2095+235−410	8	63%
4	Gemma 4 31B	2082+206−327	22	77%
5	GPT-6 Luna Decisions	2072+198−298	25	72%

Full Rotom Gen9 Bench leaderboard →

Updated 2026-10-17.

Key findings

Gemini 4 Argon currently leads the overall ranking with a rating of 2296 (95% CI 1927–2499).
Across 34 models, tokens spent per turn and win rate correlate at r = -0.18.
Ling 3.1 Flash spends the most tokens per decision: 9,210, or 6.9× the field median.
10 matches in our pilot study were won by models that produced zero tokens. These results have been retracted.
The most common signature move is earthquake, preferred by 20 models.
Mistral Large 4 attempts actions that do not exist in 3.7% of its turns, for example MOVE: earthquake.

We also report confidence intervals for Chimps at the Zoo, a bot that picks a legal action uniformly at random. It has one.