On 2026-11-05 Dwarkesh Patel released Clocktower Bench: ~$20k API for 386 social-deception games, open harness. Author board: GPT-6 Astra Elo 1859, ~82% win rate vs Opus 5.5.
Key Takeaways
- ✓Author: Dwarkesh Patel (@dwarkesh_sp) on 2026-11-05; clocktowerbench.com + Substack
- ✓Scale: ~$20k API, 386 rated games
- ✓Top: GPT-6 Astra Elo 1859; ~82% win rate vs Opus 5.5; ~+430 Elo vs Opus 4.6
- ✓Open source: github.com/dwarkeshsp/clocktower; HF dwarkesh/clocktower-games; human play
- ✓Social-deception eval is not a general IQ board; script/role variance applies
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
Public boards mostly reward correct answers, not social deception or multi-agent scheming. On 2026-11-05 Dwarkesh Patel released Clocktower Bench: models play Blood on the Clocktower, with an open harness and a human-playable site.
Architecture Highlights & Internals
One model takes all good seats, another the Demon’s team, then sides swap; a Bradley–Terry fit yields combined/detection/deception Elo. Programmatic Storyteller; private whispers and belief probes. Code and HF dataset are public; the site hosts replays and human-vs-AI play.
Authoritative Benchmarks & Measured Scores
Author figures: ~$20k API, 386 rated games. GPT-6 Astra combined Elo 1859; ~82% win rate vs Opus 5.5 (1549); ~+430 Elo vs Opus 4.6. OpenAI’s edge is mostly detection. Not a SWE-style board.
Developer Hands-on Guide
Essay: dwarkesh.com/p/clocktower; leaderboard/Play: clocktowerbench.com. Local runs need API keys. Useful for multi-agent coordination—not a general IQ total.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.