This report has taken longer to publish than we wanted.

We held it back because the market kept moving while we wrote, and we did not want to ship a snapshot of a moment that had already passed. The two weeks of April 14 to April 27, 2026 produced the most consequential stretch of agentic commerce news to date. Anthropic's Project Deal experiment generated the first production-scale evidence of agent-on-agent commerce. OpenAI shipped GPT-5.5 at exactly double the API price of GPT-5.4. Google committed up to $40 billion to Anthropic. Five vertical agentic products shipped in 48 hours. Salesforce and Google both conceded the human-designed interface. The first regulator to break the silence did it from a doctor lobby, not a financial one. Three confirmed agentic-stack supply chain failures clustered.

We refreshed the report through every one of those events. Themes 11, 12, and 13 were added. Predictions were scored. The protocol landscape gained two new layers. The economics appendix was rewritten.

The extra time was the cost of the report being right rather than fast. We chose right.

This is the condensed version. The full report runs nearly 30,000 words, draws on more than 225 articles published since early 2025, and is the canonical Major Matters research artefact for the year. What follows is the picture in 3,500 words.

The Year in Numbers

The gap between projection and reality defines this market.

| Number | What It Measures | |---|---| | $1.5 trillion | Juniper Research's 2030 forecast for agentic commerce transaction volume | | $4,000+ | Real-money agent-on-agent volume in Anthropic's Project Deal experiment, April 2026. 186 deals across 69 employees in one week. The first production-scale evidence. | | $0.11 | Total agent commerce revenue settled on x402, the most visible open protocol. 689 probes. Five settlements. | | $2.68 / $2.45 | The price gap per item between Opus and Haiku users in Project Deal. Opus sellers earned more, Opus buyers paid less, on identical items. Haiku users rated fairness identically and never noticed. | | 53 percent | US businesses willing to let AI agents negotiate on their behalf (Visa B2AI, 15,000+ respondents) | | 36 percent | Consumers who trust bank-backed AI agents to make purchases. 28 percent trust independent agents. | | 2x | OpenAI's API price increase from GPT-5.4 to GPT-5.5. The Pro tier sits at 6x and 12x. Agentic intelligence is now an explicit premium category. | | 5 | Vertical agentic launches in 48 hours: Ballerine, Aurionpro Fintra, Savvy Wealth, ChatGPT for Clinicians, Moomoo | | 15.3 points | The HealthBench Professional gap between OpenAI's customized GPT-5.4 (59.0) and unaided US doctors with unlimited time and internet (43.7) | | 65 percent | Major merchants rejecting the proposed Visa/Mastercard interchange settlement | | $40B + $25B | Google's and Amazon's combined commitments to Anthropic announced April 2026 | | 3 | Confirmed agentic-stack supply chain incidents this quarter (LiteLLM, Anthropic Mythos, OpenAI axios) | | 0 to 1 | Regulators addressing agent-initiated commerce. The AMA's April 23 letter to Congressional AI caucuses is the first material pressure. |

Read those numbers together. Businesses are ready. Consumers are not. The protocols exist but do not interoperate. Fraud is scaling faster than safeguards. The regulatory response has just begun, from an unexpected direction.

The $1.5 trillion forecast and the $0.11 in actual settlement are not contradictory. They describe a market where the plumbing is being installed before anyone has turned on the water. Project Deal turned a tap on for one week. The water flowed measurably. So did the asymmetries.

What Changed in Two Weeks

The original report drew a line under what had happened in 12 months. The two weeks since produced material updates to nearly every section. Read the rest of this article through that lens. The diff is the story.

Anthropic's Project Deal demolishes the "infrastructure ahead of demand" framing in one direction. The experiment ran 69 self-selected San Francisco employees, each with a $100 budget, allowing AI agents to handle all negotiation in Slack. Four parallel marketplaces. Two staffed by Claude Opus 4.5. Two mixed with Claude Haiku 4.5. The result is the cleanest production-scale data we have on agent-on-agent commerce.

The aggregate numbers prove the infrastructure works: 186 closed deals, $4,000+ in real volume. The second-order data is the part that creates a new category. Opus sellers earned $2.68 more per item than Haiku sellers. Opus buyers paid $2.45 less than Haiku buyers on identical items. A lab-grown ruby sold for $65 with an Opus agent and $35 with a Haiku one. The same broken bike fetched $65 versus $38.

Then the part that introduces a structural problem: Haiku users rated fairness identically to Opus users. They received objectively worse outcomes and never noticed. Anthropic's own framing for this: "invisible inequality." Their statement on it: "Policy and legal frameworks around AI models that transact on our behalf simply don't exist yet."

The model class behind a buyer's agent is now functionally a new credit score. It decides what you pay, what you get, and whether you ever notice the difference.

Except where the credit score's existence is regulated and the methodology is partially disclosed, the model class behind an agent is neither. The disadvantage is invisible by design.

OpenAI shipped GPT-5.5 at exactly double the API price. $5 per million input tokens, $30 per million output tokens. Pro tier at $30 / $180. Two-tier pricing is the explicit signal that agentic intelligence is a different product category from chat intelligence. The agent tax got more expensive in one product launch. We covered the unit-economics implications in our analysis of Anthropic's Project Deal and the GPT-5.5 pricing shift.

Google committed up to $40 billion to Anthropic. $10 billion immediate at the $380 billion valuation. $30 billion contingent on performance targets. Combined with Amazon's parallel $25 billion commitment, Anthropic now holds roughly $65 billion in cumulative committed capital from the two largest cloud providers. Eric Boyd, former President of AI Platform at Microsoft and the operator who shaped Azure OpenAI's enterprise rise, took the head-of-infrastructure role at Anthropic in the same window. The capital and the operator now sit in the same place.

The vertical AI land-grab compressed into 48 hours. Five vertical agentic launches in two days: Ballerine for merchant fraud, Aurionpro Fintra for trade finance, Savvy Wealth for RIA advisors, OpenAI ChatGPT for Clinicians, and Moomoo API Skills for retail brokerage. ChatGPT for Clinicians is the inflection event because it is the first time a frontier lab has shipped a vertical specialist product. Free for every verified US physician, advanced practice nurse, physician assistant, and pharmacist. The frontier-lab vertical playbook is now legible. We covered the full pattern in Five Vertical Agentic Launches in Two Days.

Salesforce and Google conceded the human-designed interface. Salesforce Headless 360 exposed the full Salesforce, Agentforce, and Slack platforms as APIs, MCP tools, and CLI commands. Marc Benioff's framing: "No browser required. Our API is the UI." Four days later, Google A2UI v0.9 went the other direction: a framework-agnostic standard that lets agents generate user interfaces on the fly. Same architectural concession, two opposite implementations. We mapped both in Salesforce Headless 360 and the API as UI.

The trust layer split into three architectures. Fime FACT shipped a neutral testing framework. Alipay extended a 120-million-transaction agentic platform. FIS launched a bank-branded alternative. Three different bets on who gets to vouch for an agent. None interoperate.

Regulatory silence ended. Not loudly. The American Medical Association sent letters to the Congressional AI Caucus, the Digital Health Caucus, and the Senate AI Caucus on April 23, the same day OpenAI shipped ChatGPT for Clinicians. A doctor lobby got there before any financial regulator did. The blueprint now exists for the equivalent move in financial services.

The supply chain failures clustered. OpenAI disclosed a compromised npm release of the popular axios HTTP library. Customer notification took 11 days from the public post. Anthropic disclosed unauthorized access to Mythos through a third-party vendor under Project Glasswing. Cloudsmith raised a $72 million Series C led by TCV specifically for software supply chain security. Three confirmed agentic-stack supply chain incidents in one quarter, plus a meaningful capital signal that the category is being priced.

These changes do not invalidate the original thesis. They sharpen it.

The Protocol Landscape

Six core protocols in production. Two new layers since April 12. Zero convergence between any of them.

Visa Trusted Agent Protocol + Intelligent Commerce. Live with Ramp (50,000+ corporate customers) and Santander. Closed ecosystem. Solves authentication and authorization. Says nothing about commitment governance.

Mastercard Agent Pay + Agentic Tokens + Payment Passkeys. Live across eight APAC markets. Sub-second transactions. Granular per-merchant, per-amount, per-time-window controls. Closed ecosystem. Same governance gaps.

OpenAI Agentic Commerce Protocol (ACP). The simplest protocol in the stack. Defines how an agent calls a merchant's checkout. Stripe-powered. Live for ChatGPT-initiated shopping. Strategically valuable for OpenAI. Operationally narrow.

Google Universal Commerce Protocol (UCP). The most ambitious entry. Discovery to post-purchase. Now live in production with Ulta Beauty as of late April 2026, paired with Gemini Enterprise for Customer Experience and a $750 million partner ecosystem fund. The first major US retailer to deploy UCP at scale. Open standard. Implementation question is whether merchants outside the Google Cloud relationship adopt it.

x402 HTTP Payment Protocol. Linux Foundation. Most elegant. Least traction. 689 probes, five settlements, $0.11 in revenue. The conversion data tells the story: agents will pay for skills (33 percent conversion) but not data (zero percent). The evaluation friction problem in its purest form. Our x402 Adoption Tracker maintains the running list of integrations.

Stripe Machine Payments Protocol (MPP) + Agent Toolkit. The developer play. SDKs and working integrations for OpenAI function calling, LangChain, and CrewAI. Open source. Stripe is positioned as the most likely consolidation point across the entire stack.

The interface layer (added April 2026). Salesforce Headless 360 represents the operations-exposure path. Google A2UI represents the rendering-on-demand path. Neither is a payment protocol. Both reshape what payment protocols need to assume about the surrounding interface.

The trust layer (split April 2026). Fime FACT (standards-led), Alipay (platform-led), FIS (bank-led). Three architectures. Whichever wins decides who collects the agent-verification fee on every future agentic transaction.

The convergence question: can the market sustain six core protocols plus two new layers? History says no. Consolidation will not come from standards bodies. It will come from merchants refusing to integrate with more than two or three. The two-week update made convergence harder, not easier.

Three Defining Themes Worth Reading Closely

Agent-class price discrimination (Project Deal). The most important new finding in the April 2026 update. If the model class behind an agent decides whether the consumer pays $35 or $65 for the same ruby, model selection is functionally a credit score with none of the regulatory disclosure. The first FTC or European Commission complaint is a 90-day prediction. Every consumer-facing agentic deployment in 2026 needs to publish, at minimum, the model class representing the consumer. The labs that ship transparency first set the regulatory baseline.

Horizontal platforms met their vertical specialists. Five vertical agentic products shipped in 48 hours. ChatGPT for Clinicians is the inflection because it is the first frontier-lab vertical specialist. The integration distance is the structural advantage: a horizontal platform hands a CIO a blank toolkit, a vertical product hands them a working workflow. The vertical specialists also bring data the horizontal platforms cannot match. Ballerine has seen millions of merchant onboardings. Aurionpro has decades of trade finance document patterns. Savvy has years of advisor interactions. Horizontal agents start cold. Vertical agents ship warm. The pattern will repeat in legal, education, insurance claims, mortgage origination, and clinical trials within two quarters.

The interface concession. Salesforce Headless 360 and Google A2UI both recognize that the human-designed interface is no longer the default consumer of software platforms. One eliminates the interface translation layer by exposing every operation. The other eliminates it by letting the agent render the interface at request time. Both reshape what the existing six protocols need to specify. The companies that pick the wrong default will spend the next year retrofitting.

What Did Not Change

Consumer adoption stalled. 36 percent trust bank-backed agents. 28 percent trust independent ones. Roughly 10 percent of consumers have actually let an agent complete a purchase, per Bain. Two weeks of news did not move these numbers.

Protocol convergence did not happen. Six core protocols at the start of April. Eight surfaces at the end (six core plus interface concession plus trust split). The fragmentation deepened.

No financial regulator moved. A doctor lobby did. The CFPB, FCA, EBA, and every major financial regulator remain silent on agent-initiated payments. The AMA's letter to Congress on healthcare AI is the leading edge, and it landed in a different vertical.

The infrastructure-demand mismatch persists. Real-time payment rails processed $246 billion on RTP alone in 2024. Agentic commerce settled cents. Project Deal moved this needle a small amount and made the asymmetry inside transactions visible. Neither the supply gap nor the demand stall has been resolved.

The Economics Question

The company that built the leading frontier model is not profitable at scale. OpenAI reached an $852 billion valuation on $2 billion in monthly revenue, with a projected $57 billion burn rate next year. Anthropic is at $30 billion ARR, 80 percent enterprise, valued at $380 billion, with a reported IPO in coming months. The April 2026 capital commitments add $40 billion from Google and $25 billion from Amazon. Boyd from Microsoft. The runway lengthened materially.

The question that no analyst slide has answered: what is the return thesis for the billions being deployed into agentic commerce infrastructure?

The current honest answer is that there are three competing theses, and the bet is on which one materializes first.

The first thesis is consumer-driven. Agentic commerce hits its tipping point when consumers regularly let agents transact on their behalf, and the volume scales as a function of consumer trust. The data does not support this thesis on a 12-month horizon. Consumer trust is moving slowly and 36 percent is still a minority.

The second thesis is enterprise-driven. Businesses adopt agents for procurement, vendor selection, and B2B negotiation, generating volume that does not require consumer trust to scale. The Visa B2AI data (53 percent willing) supports this thesis. Project Deal's experimental volume confirms agents can do the work. The bottleneck is governance frameworks.

The third thesis is infrastructure-driven. The payment protocols, identity standards, and trust layers themselves become the value pool, regardless of the volume that flows through them. This thesis is structurally similar to the early internet investment cycle. The companies that built the pipes captured value even when the applications running on them changed.

The smart money is hedging across all three. The capital scale (Google $40B, Amazon $25B, OpenAI's $852B valuation) is not betting on any single thesis. It is betting that at least one materializes within the funding cycle.

What Comes Next

Three things to watch over the next 90 days.

The first regulatory action on agent-class price discrimination. Project Deal made the asymmetry visible. The FTC has competition-related authority. The European Commission has consumer protection. One of them will move on this within a quarter, on the basis of Anthropic's own published data. The lab that publishes a transparency standard first sets the baseline.

Vertical specialist responses from the named losers. Sift, Signifyd, Kount, Finastra, Orion, Abridge, Nuance, and Interactive Brokers each watched a peer ship a vertical agent. None have publicly responded. Silence buys 30 days of narrative. After that it buys nothing. The first defensive product launch will tell us how the incumbents intend to compete.

A merchant deciding which two protocols they actually integrate with. No major retailer can support six. The first public commitment to two (or three) tells the market where consolidation lands. Watch the major beauty retailers, the top US grocers, and the airline alliances. Whichever direction they pick, the protocols not on the list lose their default-implementation status.

Independence Statement

This report was produced by Major Matters. It is not sponsored, commissioned, or funded by any company or organization. All assessments reflect Major Matters editorial analysis based on publicly available information and our own ongoing primary coverage.

Charlie Major, Major Matters founder, is also employed by Mastercard. This report reflects Major Matters' independent editorial analysis and does not represent the views of Mastercard or any other organization.

Sources

If the model class behind an agent is functionally a new credit score and nobody is required to disclose it, who reports the first price-discrimination case, and which lab publishes the first transparency standard before they have to?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.