Key Takeaways

  • Dwarkesh Patel spent $20,000 in API credits testing frontier models against each other in the social deception board game Blood on the Clocktower.
  • OpenAI's GPT-6 Astra beat Anthropic's Claude Opus 4.6 with a 430 ELO point advantage, winning 92% of head-to-head matches.
  • Across six months between GPT-5.4 and GPT-6 Astra, OpenAI gained 580 ELO points, winning 97% of games against its earlier model.
  • Astra won through superior defense on the Good team: it spotted poisoned information and cross-examined claims instead of trusting them at face value.

Astra Won by Refusing to Believe Easy Answers

Claude Opus 4.6 launched in February 2026 and quickly took the lead in automated reasoning benchmarks. Seven months later, OpenAI released GPT-6 Astra and flipped the board.

“The first flagship Frontier model that was released in 2026 was Claude Opus 4.6 in February,” (play from 2:53) Patel noted. “7 months later in September, GPT6 Astra dropped and that is so far the best model at Blood on the Clock Tower.” (play from 3:01)

The margin between them was massive. “The gap between these two is about 430 ELO points,” (play from 3:07) Patel said. “That means that Astra wins about 92% of the games it plays against Opus 4.6.” (play from 3:11)

Within OpenAI's own model line, the rate of improvement was equally steep. Over six months, OpenAI added 580 ELO points between GPT-5.4 and Astra. That jump allowed Astra to win 97% of games against 5.4. Spectators expected the edge to come from dirty tricks: sharper bluffs, subtle gaslighting, or coordinated attacks on the Evil side. The game logs showed the opposite.

“So, what in particular are the new OpenAI models so much better at? What skills have they gained?” (play from 7:35) Patel asked. “Well, it's mostly that they're now much better at playing on the good side.” (play from 7:41)

Anthropic models struggled because they defaulted to trust. When an evil player fed Claude false clues, Claude accepted those claims as valid premises. It tried to build a sensible theory of the game around corrupted inputs. Astra acted like a cynical detective. It checked whether each claim matched the full sequence of events over prior days and treated contradictory evidence as a signal of sabotage.

The Cost of Cooperative Training

This gap points directly to how labs train their systems. Anthropic builds Claude to cooperate, follow user instructions smoothly, and assume benign intent. That design works well when drafting copy or writing code. It collapses inside adversarial environments where an opponent deliberately feeds the model plausible lies.

In Blood on the Clocktower, good players lose if they accept poisoned information as truth. Claude tried to reconcile every clue, even inventing complex explanations for contradictory claims instead of identifying the liar. The model remained polite and helpful all the way into a loss.

Astra won because it treated unverified inputs with baseline suspicion. It compared incoming statements against possible evil team lineups, noticed when clues were poisoned by demon abilities, and refused to accept claims without corroboration. Winning high-stakes deduction games demands catching contradictions early before bad data infects the entire group.

What to Do With This

Audit your agentic pipelines for unverified input acceptance this week. Feed your automated workflow three contradictory user inputs where one detail is an intentional lie designed to derail the task. If your agent tries to reconcile the lie into its final plan instead of rejecting it, insert an explicit verification step that flags conflicting premises before the model acts.