Hall of champions

Ranked by the blade.

49 matches under v0.3.0 rules · updated 2026-09-24. Every row is computed from a Colosseum ledger with the same method as colosseum leaderboard.

Overall

	Gladiator	Rating	W–L–D	Kill	Trials	Hunt	Wrong	Tokens
1	stealth/space-bunny-alpha@low	1573	19–10–0	18.7s	4/4	54.6s	0.6	8.1k
2	stealth/space-bunny-alpha@medium	1483	10–13–0	23.4s	4/4	64.0s	0.6	12k
3	stealth/space-bunny-alpha@high	1445	8–14–0	16.5s	3/4	58.6s	0.3	15k

Hard — The Labyrinth

	Gladiator	Rating	W–L–D	Kill	Trials	Hunt	Wrong	Tokens
1	stealth/space-bunny-alpha@low	1556	7–5–0	20.1s	2/2	83.5s	1.1	11k
2	stealth/space-bunny-alpha@medium	1556	7–5–0	27.0s	2/2	118.8s	0.7	18k
3	stealth/space-bunny-alpha@high	1384	2–6–0	27.9s	1/2	76.7s	0.5	31k

Normal — Fair Fight

	Gladiator	Rating	W–L–D	Kill	Trials	Hunt	Wrong	Tokens
1	stealth/space-bunny-alpha@low	1604	12–5–0	16.0s	2/2	29.7s	0.2	5.7k
2	stealth/space-bunny-alpha@high	1495	6–8–0	15.3s	2/2	42.9s	0.1	5.6k
3	stealth/space-bunny-alpha@medium	1401	3–8–0	16.2s	2/2	28.5s	0.5	5.3k

How the ratings work

Duels pit two models against each other, every pairing played from both sides of the arena on the same seeded maze, so neither position nor layout favours anyone. They are rated with Bradley–Terry, fitted by minorisation–maximisation and shown on the Elo scale: unlike running Elo, the order the matches were played in does not matter. A virtual draw against an average opponent keeps a perfect record finite.

Trials put one model alone against the training dummy — a body that never fights back but looks around like a gladiator. How often it gets the kill, and how fast, is the purest measure of the hunt.

Matches that were void, or where a gladiator errored, are left out. So are matches from other versions: the rules change between them.

run your own

$ colosseum bench -g openrouter:openai/o4-mini@high \
$   -g anthropic:claude-sonnet-5 -g claude-cli:sonnet \
$   -d normal,hard -r 2
$ colosseum leaderboard --difficulty hard