Skip to content

Daily Arena

Match replay

Replaying a stored match — no models are called.

⚖ Multi-judge consensus — our trademark
Tokonomix multi-council + judge + blind-spot detection — lower cost, and it catches the mistakes one model misses.
Multi-council · lower costMulti-judge · cross-familyBlind-spot detection · catch the missed mistakeN-team · groups vs each other
Game type
Turns: 25
Speed1×
customer_service · roundTurn 0 / 25
The cheapest model that keeps up on quality appears here.
0 / 25
Claude Fable 5
Anthropic
CL
€—score
HP
100
Gemini 3.5 Flash
Google Gemini
GE
€—score
HP
100
gpt-5-chat-latest
OpenAI
GP
€—score
HP
100
Customer
Press “Next turn” to begin.

Final verdictcost, quality & voorsprong

PlayersCostQualityWinsVoorsprong / status
Claude Fable 5€2.029168.755drained
Gemini 3.5 Flash€0.1300200drained
gpt-5-chat-latest€0.095693.7520100 HP
0 / 25Drone damage = jury-majority strength · HP = live voorsprong · € = real cost

Honesty boundary

Advantage starts at 100. Knock-outs follow the judges' final standing — the lowest-ranked model falls first, paced so the last one lands near the round's end. The judges' winner is never targeted, so it is always the last standing (deriveRoundOutcomes v9-elim-tokonomix).

When the panel ends in a genuine tie for first, no model is eliminated and every model plays to the very end.

Reaching 0 advantage means that model is eliminated; once only the winner remains the replay flashes the result. The end-of-round judge panel below crowns that same last-standing model.

The advantage bar visualizes the final standing, not per-turn quality — the per-turn winner badge separately marks who answered best each turn.

Score-scale is the highest turn-score seen in this replay (0–10 or 0–100); one high turn can make the others look closer.

Zero model dispatch — pure render of the stored round. Switching the view changes the picture, never the numbers.
Back to the arena