@readmodelmag
MODEL
Subscribe

AI News

He spent $20,000 so AIs could play Werewolf

SummaryDwarkesh Patel's Clocktower Bench (Nov 5) ran 386 Blood on the Clocktower games, model vs model. GPT-6 Astra leads Claude Opus 4.6 by ~430 Elo (~92% wins). The edge is detection, not better lying.

· MODEL Editors · 1 min read

Some links may be affiliate links. Disclosure

Cartoon split panel titled He spent $20,000 so AIs could play Werewolf. The winner wasn't the best liar. It was the best detective. Left EVIL TEAM: three scheming generic robots whisper psst and bluff over a $20,000 API receipt for 386 games. Right GOOD TEAM: blue detective robot with magnifying glass catches the Demon; score card shows GPT-6 Astra ~430 Elo over Opus 4.6, edge on detection not lying. Watermark MODEL · readmodelmag.com. No real faces or logos.
Our cartoon. $20k of Werewolf. Best detective wins. Illustration by MODEL. Original. Image source.

Dwarkesh Patel spent about $20,000 of API credits so AIs could play Blood on the Clocktower — social deception like Mafia/Werewolf, Demon + evil vs good town. On November 5, 2026, he published Clocktower Bench.

The setup. 386 games where the whole good team is one model and the whole evil team is another. Models have to lie, bluff, and catch liars. Site: clocktowerbench.com (opens in a new tab).

Who's winning. GPT-6 Astra (OpenAI, Sept 2026) is so far the best. The gap vs Claude Opus 4.6 (Feb 2026) is about 430 Elo — Astra wins roughly 92% of those matchups. Dwarkesh's own X follow-up puts Astra at an ~82% win rate against Opus 5.5. Over about six months from GPT-5.4 to Astra, OpenAI gained roughly 580 Elo (~97% win vs 5.4).

The punchline that is also the result. OpenAI's lead is mostly on the good/detection side — catching liars — not better lying. Anthropic and OpenAI are roughly level as deceivers. The winner of the $20k Werewolf tournament wasn't the best liar. It was the best detective.

Caveats. This is one hobby benchmark, not a general model ranking. Numbers and claims are from Dwarkesh's essay and follow-up. Not tested by MODEL. Not sponsored.

Sources: Dwarkesh essay (opens in a new tab); clocktowerbench.com (opens in a new tab); Dwarkesh X (opens in a new tab).

MODEL Editors

Sources

  1. The benchmark where AIs have to lie to win (Nov 5, 2026)

    Dwarkesh Patel

  2. Clocktower Bench

    clocktowerbench.com

  3. Dwarkesh X follow-up on Astra vs Opus 5.5

    Dwarkesh Patel on X

Keep reading

  • MODEL editorial cartoon on cream. Left door: ANTHROPIC with red IPO banner and sticky note GONE on empty Claude Code Product desk. Stick figure walks right carrying a SHIPPED box toward an open OPENAI Core Products door. Caption: Claude Code's product lead just shipped herself to OpenAI. Watermark MODEL · readmodelmag.com.

    AI News

    Claude Code's product lead just shipped herself to OpenAI

    Cat Wu (@_catwu), former Head of Product for Claude Code (and Cowork) at Anthropic, is joining OpenAI to lead product management for Core Products — weeks before Anthropic's planned IPO. OpenAI confirmed to Business Insider; Anthropic declined comment.

    · MODEL Editors

  • MODEL news card on a cream background. Small red kicker: AI NEWS. Large red figure: 99%. Headline: Status line cut its own bill. MODEL wordmark bottom left; Source: ClaudeDevs bottom right.

    AI News

    Claude Code cut the status-line tax by ~99%

    Background subagent status used to re-read the whole chat every 30 seconds. Now Haiku 5.5 writes it from short notes — Anthropic says each update costs about 99% less, and sessions with background subagents used ~5% less of the usage limit (~6% more prompts). CLI v2.1.316+.

    · MODEL Editors

  • MODEL news card on a cream background. Small red kicker: AI NEWS. Large red figure: READ. Headline: Claude can read your bank (read-only). MODEL wordmark bottom left; Sources: Anthropic, Plaid bottom right.

    AI News

    Claude can now read your bank accounts (read-only)

    On Nov 5, 2026 Anthropic launched Personal Finance in Claude with Plaid: connect US bank and card accounts, ask in natural language, read-only access. Beta this week for Pro, Max, and Ultra in the US — gradual rollout, paid plan does not mean instant access.

    · MODEL Editors

FREE · EVERY WEEK

The top AI tools this week, tested.

Every week MODEL's editors pick the AI tools worth your time, try them hands-on, and tell you what's good, what's not, and who it's for. Free, in your inbox.

Free. Unsubscribe anytime. Our picks are editorial and can't be bought.