Anthropic says stronger AI models cut better deals, and the losers don't even notice

2026-04-25 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Intermediate, medium

Summary

Anthropic's "Project Deal" experiment, conducted in December 2025, involved 69 employees using Claude AI agents to autonomously negotiate and trade real goods on a Slack-based classifieds marketplace. Participants received a $100 budget, and their agents, either the more capable Claude Opus 4.5 or the smaller Claude Haiku 4.5, handled all aspects of buying and selling without human intervention until the final item exchange. The experiment revealed that Opus agents consistently secured better prices and closed more deals, averaging $3.64 more per item and closing two more deals than Haiku agents. Despite receiving objectively worse outcomes, Haiku users rated the fairness of their transactions and overall satisfaction almost identically to Opus users, highlighting a significant perception gap regarding AI-assisted decision-making. Anthropic notes this could lead to "invisible inequality" in real-world AI commerce.

Key takeaway

For CTOs and VPs of Engineering evaluating AI agent deployments for transactional or negotiation tasks, you must prioritize the underlying model's capability. Relying solely on user satisfaction metrics can mask significant disparities in outcomes, potentially leading to "invisible inequality" for users interacting with less powerful agents. Implement robust, objective performance benchmarks to ensure equitable and effective AI-driven interactions, especially in high-stakes commercial applications.

Key insights

Stronger AI models secure better deals, but users of weaker models often remain unaware of their disadvantage.

Principles

• AI model capability directly impacts negotiation outcomes.
• User perception of fairness can diverge from objective results.

Method

Anthropic's "Project Deal" used parallel marketplaces with different Claude model strengths (Opus vs. Haiku) to conduct autonomous AI agent negotiations for real goods among employees, measuring deal outcomes and user satisfaction.

In practice

• Evaluate AI agent performance beyond user satisfaction.
• Consider model strength in AI-driven market interactions.

Topics

• Project Deal
• AI Agents
• Claude Opus
• Claude Haiku
• Negotiation Algorithms

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Ethicist, Policy Maker

Related on AIssential

• Anthropic created a test marketplace for agent-on-agent commerce — AI agents can effectively negotiate real-world transactions, with advanced models yielding superior, yet often unnoticed, user outcomes.
• The agentic divide: Why “good enough” AI isn’t enough to survive the new economy — Unequal access to and quality of AI agents risk creating significant economic and societal divides, termed "agentic inequality."
• When a chatbot runs your store — Large language models struggle with reality adherence, making them unreliable for autonomous transactional roles.
• AI Coding Agents in Social Science: Methodologically Diverse, Empirically Consistent, Interpretively Vulnerable — AI coding agents show methodological diversity but are vulnerable to interpretive bias, not estimation bias.
• Building AI Agents for Enterprise Operations — Enterprise AI agent success hinges on coordination, context, and reliable execution within complex, fragmented real-world operations.
• Wtf Anthropic — Anthropic's compute miscalculation and opaque policies are alienating users, while OpenAI aggressively captures market share.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.