Anthropic's Project Deal: Claude agents ran a marketplace

Quick Summary

Anthropic published Project Deal on April 24, 2026: 69 employees, 186 deals, $4,000+ transacted, every transaction mediated by Claude agents on both sides
Participants represented by Opus 4.5 systematically out-negotiated participants represented by Haiku 4.5 - first quantified evidence of agent-vs-agent asymmetry
Buyer-side agents are already shipping (ChatGPT, Gemini, Comet); seller-side agents lag, putting undefended catalogs at a structural negotiating disadvantage
Sellers should audit agent-readable structured attributes, lengthen attribution windows, tag agent-referral traffic, and set hard pricing floors at gross-margin thresholds

Nova surfaces every Amazon fee, refund, and margin shift in your live P&L, across 21 marketplaces. Explore the live P&L

What's happening

On April 24, 2026, Anthropic published the results of Project Deal, a one-week internal marketplace experiment in which Claude agents bought and sold real items on behalf of 69 Anthropic employees in its San Francisco office. Each participant got $100 and was paired with a Claude agent on either side of every transaction. The agents struck 186 deals worth roughly $4,000. From the agency portfolios we monitor, the cohorts most exposed feel it inside a billing cycle, not a quarter.

The headline finding, summarized by Unite.AI and PYMNTS, is that the strength of the model on each side measurably changed the outcome. Participants represented by Opus 4.5 systematically walked away with better prices than participants represented by Haiku 4.5, and the human counterparties never noticed. This is the first published experiment that puts a number on agent-vs-agent commerce asymmetry.

Project Deal is small in dollar terms but large in implication. Digital Commerce 360 Read it as a signal that Anthropic is now actively investing in commerce surfaces alongside OpenAI and Google. TechSifted Framed it more bluntly: when both sides of a transaction are AI agents, the human shopper's price-sensitivity, brand affinity, and impulse behavior stop mattering. The deal terms become a function of model quality and prompt design.

Deals closed

186

Across 69 Anthropic employees in one week

Real money transacted

$4,000+

Per Anthropic's published results

Model gap

Opus > Haiku

Stronger model = measurably better terms

Key Dates & Deadlines

Apr 21, 2026

Project Deal week begins

Anthropic gives 69 employees $100 each and pairs every buyer and seller with a Claude agent in a private internal marketplace

Apr 24, 2026

Anthropic publishes Project Deal results

186 deals, $4,000+ transacted, agents on Opus 4.5 systematically out-negotiated agents on Haiku 4.5

Apr 27, 2026

Industry coverage lands

PYMNTS, Unite.AI, TechSifted, and Digital Commerce 360 publish analyses framing it as the first hard data on agent-to-agent commerce

Why Amazon and multi-channel sellers should care

Buyer-side agents are coming faster than seller-side ones

Project Deal is internal, but it sits on top of a public stack already shipping. ChatGPT checkout, Google Gemini's agentic checkout with Shopify and Walmart, and Amazon's Perplexity Comet integration all give buyers an agent. Most sellers do not have a counterparty agent yet. Until they do, Anthropic's data suggests they will systematically lose price negotiations to better-equipped buyer agents.

The Buy Box logic is built for humans, not agents

Today's Buy Box, deal pages, and search ranking weight signals (rating, brand familiarity, image quality, badge density) that an agent buyer can flatten. An agent comparing 12 ASINs by structured attributes does not care about your hero image. It cares about price, restock probability, return rate, and structured spec match. That changes which catalog edits actually move conversions and which are visual debt.

Last-click attribution gets worse, not better

When a buyer agent reviews 50 SKUs, picks one, and clicks through to checkout, the last touch is the agent referral. Whatever Sponsored Products, Sponsored Brands, or DSP impression seeded that agent's preference is now unattributable in standard reports. This compounds the attribution erosion already documented around Rufus, agentic checkout, and AI Mode in Marketplace Pulse's recent commentary.

What you should do now

1.
Audit which SKUs are agent-readable
Review structured attributes (bullets, A+ feature blocks, GS1 attributes) for the 20 SKUs that drive 80% of revenue. Agents read structured fields before they read images. Baymard Institute's product-page research is a good operator baseline.
2.
Lengthen attribution windows
Agent-mediated purchases happen further from the original ad impression. Move PPC efficiency reviews to a 30-day minimum window and reconcile against actual P&L, not against the last-click ACoS the ad console reports.
3.
Track agent referral traffic separately
Tag traffic from ChatGPT, Perplexity, Comet, Gemini, and Claude as its own channel. Even if conversions are small today, the cohort behavior and basket composition are different. Practical Ecommerce has covered the early implementation patterns.
4.
Pressure-test your pricing engine
If Anthropic's data holds at scale, an undefended catalog will be repriced down by buyer agents that compare across marketplaces in real time. Set hard floors at gross-margin thresholds, not at competitive percentile thresholds.

How Nova helps on the marketplaces it covers

Nova does not build buyer agents and does not integrate with Anthropic, Project Deal, or any agent-marketplace. It covers the data layer for Amazon (SP-API across 21 marketplaces) and Walmart, where the practical impact of agent-mediated buying first shows up in attribution and margin reports.

For attribution that survives longer purchase paths, the relevant surfaces are PPC Analytics with extended windows and Profit and Loss Reconciled at SKU level. To re-baseline what drives conversion, Custom Analytics and the analysis-ready data feed let teams query agent-referral cohorts without rebuilding the pipeline.

Agencies and aggregators running portfolio-level reads can review the agency workflow and the aggregator workflow for how multi-account teams structure attribution audits. Strategy context lives in multi-marketplace analytics and Amazon vs Walmart 2026.

Ready to Transform Your Amazon Business?

Join thousands of successful sellers who use Nova Analytics to make data-driven decisions and maximize their profits.

Try For Free Book Demo

Get More Amazon Seller Tips

Subscribe to our newsletter for weekly insights, strategies, and market updates.

Frequently Asked Questions

Common questions about this topic

Project Deal was a one-week internal marketplace experiment Anthropic published on April 24, 2026. Sixty-nine employees in its San Francisco office got $100 each and were paired with Claude agents on both sides of every transaction. The agents struck 186 deals worth roughly $4,000 in real goods.

It produced the first quantified evidence that the model strength on each side of a transaction measurably changes the outcome. Participants represented by Opus 4.5 systematically walked away with better prices than participants represented by Haiku 4.5, and the human counterparties never noticed.

Buyer-side agents are already shipping in ChatGPT, Gemini and Perplexity Comet. Seller-side agents lag. If Anthropic's data holds at scale, undefended catalogs will be repriced down by buyer agents that compare across marketplaces in real time. Sellers should set hard pricing floors at gross-margin thresholds, lengthen attribution windows, and audit agent-readable structured attributes.

No. Nova does not build buyer agents and does not integrate with Anthropic, Project Deal, or any agent-marketplace. It covers the data layer for Amazon (SP-API across 21 marketplaces) and Walmart, where the practical impact of agent-mediated buying first shows up in attribution and margin reports.

Verified Sources

All information verified from official Amazon sources and trusted industry analysts as of publication date.

Never Miss a Critical Amazon Update

Get breaking news, policy changes, and time-sensitive updates delivered to your inbox.

Deep Dive: Related Guides

For more comprehensive analysis on these topics:

→ Amazon Rufus Impact on Sellers: AI Analytics Strategy 2026

Amazon Rufus changes how buyers shop. Learn how to track AI-driven behavior shifts, optimize for conversational search, and adapt your analytics strategy.

→ Amazon Multi-Marketplace Analytics

Your UK marketplace might be 18% more profitable than US, but Seller Central won't tell you. Learn how to consolidate analytics across Amazon marketplaces with currency normalization and marketplace-specific cost tracking.

→ Amazon Cross-Marketplace Reporting

Managing sales in US, EU, and Japan means three separate Seller Central accounts. Without unified reporting, you can't see which marketplace is actually profitable. This guide shows how to consolidate data, normalize currencies, and build a global P&L.