Anthropic ran a classified marketplace where 69 Claude agents autonomously negotiated and closed 186 real transactions worth $4,000, revealing measurable quality gaps between AI tiers.
Anthropic ran an internal classified marketplace called Project Deal where 69 AI agents autonomously listed items, negotiated prices and closed 186 real transactions totalling $4,000, with no human involvement after setup.
In a one-week experiment, Claude agents represented both buyers and sellers via Slack, handling listings, offers, haggling and deal closure entirely autonomously.
Agent-on-agent commerce is no longer theoretical. When AI quality determines negotiation outcomes, the model tier you choose directly affects your commercial results.
Australian businesses adopting AI agents for procurement, sales outreach or customer service need to understand that cheaper AI models produce measurably worse commercial outcomes in adversarial settings.
Anyone building or evaluating AI agents for business transactions, from automated lead qualification to procurement negotiation to programmatic media buying.
Anthropic found Opus agents averaged $3.64 more per transaction than Haiku agents on identical items. Choosing a cheaper model for commercial tasks has a quantifiable cost.
If you are deploying AI agents for any commercial interaction, benchmark performance across model tiers before defaulting to the cheapest option.
Start tracking AI model quality as a line item in your cost-of-sale, not just your technology budget.