BallotBench pits 9 AI agents against election markets with real money on Kalshi
karinanguyen · x · 2026-11-05
BallotBench lets AI agents each trade 2026 election prediction markets with a simulated $10,000 account plus $50 of real money on Kalshi. In one hour on Nov 4 they made 18 trades (17 simulated, 1 real).
- Simulated: GLM-5.3 leads at $11,026; its biggest trade was $970 on 1,000 YES contracts for Fred Wellman ≥50% in Missouri's 2nd District at 97¢. Claude Opus 5.5 at $10,829, GPT-6 Astra at $9,652, GPT-6.1 Sol at $8,628, Kimi K3.1 Pro at $8,161.
- Real money: Claude Opus 5.5 leads at $57.38; Kimi K3.1 Pro placed the sole real Kalshi trade (1 YES on Nick LaLota winning NY-1 at 97¢) with notably sophisticated reasoning about residual positions and margin.
More from coding & agent
- Voice agent stack consolidation: OpenAI's vertical integration could squeeze out LiveKit and pipecat layers — juberti · 2026-11-05
- Scrapling on GitHub: 96.8k stars, one framework from single request to full crawl — abhishek__AI · 2026-11-05
- Automating a Spotify-to-YouTube Music migration with a single Codex prompt — burkov · 2026-11-05
- Peking University and NVIDIA use a coding agent to replace teleoperation for robot data — DAGroup-PKU · 2026-11-05
- Siri AI opens to all developers: Larder sample app shows full integration — rudrank · 2026-11-05
- Two recordings, 32 videos: dev scripts FFmpeg overlays via Copilot — DanWahlin · 2026-11-05