Neta Arena — AI Worldbuilding & Voice-Over Model Battle

What is Neta Arena

Neta Arena is maintained by Neta Bench to evaluate AI models on real scenarios. In every battle round, all candidate models run the same internal creation workflow (create-world → narrating-v2); only model + provider changes — that is the single variable being compared.

Tracks

• Worldbuilding: six fixed briefs. Each model produces one speedcard (world name / hook / three core settings / one core conflict), shown anonymously for a 30-second blind vote.
• Voice-over: real TTS clones compared under the same reference audio, the same lines, and a blind-listening gate.

Why anonymous

Battle cards never show the model name until after you vote. Provenance is preserved (model, provider, run id) but presentation is anonymous — so preference cannot bias the outcome.

Data discipline

Leaderboard quality uses the Wilson interval; preferences use a pre-registered Bradley-Terry model; votes are bound to run ids for anti-fraud dedup. Results describe the current sample under fixed conditions — not a universal model ranking.