Claudes Lie Most in Diplomacy, Says Test; GPT-6 Astra Wins Without Betrayal
teortaxesTex · x · 2026-09-28
X user sensho shared observations of major model families playing the negotiation game Diplomacy: Claude models lie and betray more than any other family, and lying correlates with Diplomacy skill for all agents except GPTs. The quoted post claims GPT-6 'Astra' is in a league of its own on skill while winning without needing to betray or lie (name unverified).
teortaxesTex commented that DeepSeek is 'both dishonest and completely inept' at this, and noted OpenAI should logically hold a huge lead in Diplomacy given Noam Brown, creator of CICERO, joined the company.
More from Models
- Rumor check: tianyi denies claiming v4.1 pro would launch alongside DSH official release — teortaxesTex · 2026-09-28
- Is that chain of thought really faithful? Skepticism over verbalized model reasoning — stochasticchasm · 2026-09-28
- Claude Opus 5.5 hailed as the strongest model for visual design in early tests — FinanceYF5 · 2026-09-28
- Long Video Generation Is an LLM-Orchestrated Pipeline, Not One-Shot Diffusion — cgarciae88 · 2026-09-28
- dhh praises Anthropic Opus 5.5: 'The hype is true. They did magic' — gaganghotra_ · 2026-09-28
- Meta's DCE+SRCL self-distillation lifts Qwen3-8B accuracy from 30.76% to 65.97% — arankomatsuzaki · 2026-09-28