AI BriefingPreprint: Frontier LLM Agents Fail to Invest in Renegotiable Option Games
·2026.11.03 12:17
Key point
A new preprint finds that Claude Opus 5.5 and GPT-6.1 Sol never invested in renegotiable option scenarios, contrasting with a 2011 human study where such options raised investment to 67.2%.
Details
- A new preprint (arXiv 2610.40961) demonstrates that two profit-maximizing frontier LLM agents, Claude Opus 5.5 and GPT-6.1 Sol, failed to invest in a classic economics hold-up game when renegotiable options were available.
- In 1,440 one-shot games, sellers adapted in 100% of non-renegotiable option scenarios (288 games) but in 0% of scenarios involving no contract, non-binding reference prices, or renegotiable options (864 games combined).
- The study contrasts these AI results with a 2011 human lab study by Hoppe and Schmitz, where renegotiable options increased the share of seller adaptation from 41.5% (22 of 53) to 67.2% (43 of 64).
- Robustness checks showed that removing the "maximize profit" instruction did not change seller choices, and sellers did not invest when adaptation only added cost, confirming the agents' strict adherence to profit maximization over strategic cooperation.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Y Combinator Labs and Olam Labs run multi-agent AI swarms in Minecraft and MindustrySemiAnalysis: Anthropic Subscriptions Offer 5x More Value Than OpenAI
Agent Swarm Experiment: Model Choice and Prompt Order Drive Emergent LeadershipCognition releases FrontierCode 2.0, a coding benchmark where top models score only 45%