A procurement officer sits down on a Monday morning to build a budget for an enterprise AI agent rollout. They open eight browser tabs, one for each of the primary platforms: OpenAI, Anthropic, Google, Microsoft, xAI, Meta, ServiceNow, and Salesforce. The goal is simple — find a per-seat or per-unit price to drop into a spreadsheet. By the time they reach the end of the list, no single metric has appeared that lets them compare one agent against another. What they have instead is a row of quote-gated portals, usage meters and third-party estimates: enough material to fill the spreadsheet, not enough to defend it.
Four of the eight publish no public enterprise-agent price on their own surface. OpenAI Presence lists no price on its product page or its announcement; the service is not self-serve, is limited to eligible enterprise customers, and routes buyers to contact sales or an account team — a quote-gated enterprise SKU, where the pricing mechanism exists and simply is not shown. Meta’s Business Agent announcement says getting started is free and that paid subscription offerings are planned, but publishes no seat or subscription price. Meta’s own developer documentation — which returned a 400 on direct fetch and reached us through an indexed snippet plus two independent third-party guides — carries the one Meta number that exists: $2.00 per 1 million tokens for Business Agent replies on the WhatsApp Business Platform, effective August 1, roughly four to five cents a message. The enterprise custom-agent platform stays unpublished. ServiceNow’s AI agents page carries a “Contact us for a quote” call to action and the site has no official pricing page at all; third-party estimates put ITSM at $70 to $200+ per fulfiller per month, which are not ServiceNow figures. Google Gemini Enterprise closes the four: its official pricing surfaces display no per-seat dollar rate, and the $21-to-$60 range in circulation is third-party reported US list pricing, never confirmed on a current Google page.
The two that do publish hand you units that refuse to line up. Microsoft Copilot Studio sells a capacity pack of 25,000 Copilot Credits at $200.00 a month off a published rate card, alongside a pay-as-you-go meter through Azure. Microsoft has moved off the older model in which 25,000 messages cost $200, and because consumption now depends on how an agent is designed and used, a message no longer maps to a flat price. Salesforce Agentforce, whose pricing page was last modified October 19, publishes the fullest card in the group — Flex Credits at $500 per 100,000 credits with each Agentforce action drawing 20 of them, a $5 per-user license, a $125 add-on, a $150 industry add-on, a $50 Coworker license, a Max Edition from $550 — all of it off that page. It also shows $2 per conversation, except the card carrying that price now reads, to anyone arriving as a new buyer, “For existing Agentforce Conversations customers. Contact your Account Executive for more information.” A published price, retired inside the page that still prints it. Credits, conversations, actions and seats do not convert into one another, which is why two vendors with honest rate cards still cannot be ranked against each other from public information — the problem four rival agents and the infrastructure pricing gap both already set out.
Anthropic and xAI sit in between. Anthropic’s Enterprise plan lists $20 per seat per month billed annually — a figure present in Anthropic’s own pricing page markup while its official support article describes the structure without restating it — and usage is not included: Claude chat, Claude Code and Cowork bill separately at standard API rates. xAI quotes enterprise deals and publishes no enterprise rate card, though its standard API card is public at $2.00 input and $6.00 output per million tokens for grok-4.7, and, as we reported in October, GrokBot’s seat price is public while the deployment bill behind it is not.
On December 1, 2026 — five weeks out — Microsoft changes three things at once, per its Partner Center October announcements, a page whose own metadata shows it was last updated October 14. New Microsoft 365 Copilot and Copilot Business offers can be bought without a qualifying Microsoft 365 base subscription, across all channels including CSP, priced the same as the current offers on new product and SKU IDs. Offers that still require a base subscription pick up “(add-on)” in their names from October 16 and reach end of sale in CSP on December 1 — end of sale, in Microsoft’s own wording, is not end of life: existing customers can keep renewing and adjusting seats, and nothing on existing subscriptions changes. And usage-based billing becomes the default billing configuration for new Microsoft 365 Copilot Business subscriptions bought through CSP, a date moved back from November 2. The Azure subscription setup arrives with the license, and an adjustable default spending limit of 4,000 Copilot Credits per user per month comes pre-set — Microsoft’s own example has 100 users consuming up to 400,000 credits a month. The experiences it names as usage-based are Copilot Cowork, Work IQ APIs and GitHub Copilot Harness; Microsoft’s usage-based billing documentation adds that core per-user Copilot in Word, Excel, Outlook, Teams and Chat stays under the fixed license. The US is not on the initial exclusion list.
Set the two facts beside each other and the problem lands on a person. A procurement team that cannot see a number on four of eight platforms now also watches the one vendor that does publish numbers turn its own entry point into a variable — on a default nobody chose. Their reference rollout is 30,000 seats at an undisclosed cost, measured on adoption rather than productivity. Underneath sits the labor nobody has priced: human review is load-bearing and priced at zero, and a meter running by default does nothing to reveal it.
The narrow claim, then: not that anyone is hiding anything, but that there is no common unit to compare them with. Until one gets published, the cost of an agent is something you find out after the contract is signed, not before.
