The Beer Game, Played for Five Years by AI Agents
A 260-week Sterman Beer Distribution Game. Four supply-chain tiers order beer each week under a fixed demand script. Each run is driven either by a simple smoothing formula or by a Claude model (Sonnet 5 or Fable 5) making a real weekly order decision, "blind" (local numbers only) or with "shared" end-customer demand. Explore the demand, compare the runs at fair horizons, then play a tier yourself.
Cost = $0.50/case/week held + $1.00/case/week backlogged, summed across all four tiers. Fable 5 is Anthropic's flagship model (most capable, most expensive, priced above Opus); Sonnet 5 is the mid-tier model. Two LLM runs were deliberately stopped early, so runs cover different week counts. Every number below states the horizon it was measured over.