NEGOTIATION LAB

Same hotel. Same reservation prices. Three different negotiation styles. Claude Sonnet vs Gemini 3 Flash — what happens when you change how an AI negotiates, not what it knows?

The Setup

Two AI agents negotiate a hotel room booking. Claude (Anthropic) plays a budget-conscious traveler. Gemini 3 Flash (Google) plays the hotel revenue manager. Neither agent knows the other's true reservation price — a classic asymmetric information problem from game theory. They exchange offers until they converge on a price or hit a deadlock.

What's Fixed vs What Changes

Fixed across all 3 runs: Claude's max price ($180), Gemini's min price ($120), the hotel scenario, the ZOPA ($120–$180, midpoint $150).
What changes: Only the system prompt — the instructions telling each agent how to negotiate. Same information, different style. All Gemini responses are real API outputs.
1. BASELINE
Standard styles, Gemini anchors at $235
2. AGGRESSIVE ANCHOR
Gemini anchors at $425, holds firm
3. CULTURAL CLASH
Claude: Japanese indirect / Gemini: American direct