Agentforce vs. Copilot Studio vs. Gemini Enterprise vs. OpenAI Agents API vs. Claude Managed Agents: Five Billing Units for One Enterprise Agent

The decision rule

Buy the platform whose billing unit you can measure in your own telemetry before you sign.

If you can count actions, Salesforce's Flex Credits are legible. If you can count model turns per conversation, Microsoft's credits are legible. If you can count tokens and minutes of runtime, Anthropic and OpenAI are legible. If you can only count employees, Google's seat model is legible. A unit you cannot forecast from your own data is a unit you will be surprised by in month four, and on one of these platforms the surprise is your agents being switched off.

The five units, defined from the pricing pages

Prices and plan names are as published on each vendor's own pricing page; last verified 3 October 2026. Microsoft's figures are from its product page and its Learn billing page.

Microsoft's 25 September 2026 announcement matters here too: the per-user Copilot licence covers "Copilot in Chat and across Word, Excel, PowerPoint, Outlook and Teams," while "Cowork, Code, and Autopilot, new long-running agentic capabilities, and frontier models like Astra and Fable all run on UBB," usage-based billing. Long-running agents are being moved off the seat everywhere; Google is the holdout, and only because its price list is private.

The same agent on each platform

The agent: a customer-facing support agent handling 900 conversations a day, 27,000 a month. Assume each conversation is four model turns, one of which calls a tool (order lookup). Where tokens are the unit, assume 3,000 input tokens per conversation (2,000 cached, 1,000 fresh) and 700 output. Where runtime is the unit, assume four minutes of active running per conversation.

Microsoft Copilot Studio. Microsoft's own example on the billing page is almost this agent: "[(4x1)+(2x2)] x 900 customers = 7200 Copilot Credits" a day, for four classic answers and two generative answers. That is 216,000 credits a month, or 8.64 packs, so nine packs at $200 = $1,800 a month. A more realistic generative agent (three generative answers at 2 each plus one agent action at 5, so 11 credits per conversation) is 9,900 a day, 297,000 a month, twelve packs, $2,400 a month. Buy nine packs and run the second pattern and you hit 132% of capacity, past the 125% line: your agent is disabled mid-month unless a pay-as-you-go meter is linked. Link the meter.

Salesforce Agentforce. On conversations: 27,000 × $2 = $54,000 a month. On Flex Credits with four actions per conversation: 80 credits = $0.40 per conversation, so $10,800 a month. The two Salesforce units cross at twenty actions per conversation ($2 ÷ $0.10). Below that, credits are cheaper; above it, the flat conversation price is. A simple support agent is well below twenty, so the $2 conversation price is for long, action-heavy conversations and voice. Add the $5 per user licence for any employee who needs to interact with the agent's records, and note that conversation pricing is Pre-Purchase only, so you are forecasting volume up front.

Anthropic Claude Managed Agents, Sonnet 5.5. Tokens: 1,000 × $2 + 2,000 × $0.20 + 700 × $10, per million, is $0.0094 a conversation. Runtime: 4 ÷ 60 hours × $0.08 = $0.0053. Total about $0.0147, so about $400 a month. On Opus 5.5 the token line rises to $0.0184 and the total to about $0.0237, so about $640 a month. Runtime is roughly a third of the Sonnet bill, which is why "idle is free" is not a footnote: an agent that waits on humans costs tokens, not hours.

OpenAI Agents API, GPT-6.1 Sol. Tokens: 1,000 × $2 + 2,000 × $0.10 + 700 × $10, per million, is $0.0092 a conversation, so about $250 a month in model cost. The swing is the container. The published unit is "per 20-minute session per container," and the same page says sessions are "billed by the minute, with a 5-minute minimum per session." So a short support conversation that opens its own 1 GB container pays the minimum: 5/20 × $0.03 = $0.0075, and 27,000 of them are about $202 a month on top; only a conversation that holds the container for the full 20 minutes pays $0.03 ($810 a month across the lot). If sessions are pooled or the agent needs no sandbox, it is near zero. Your architecture decides whether this platform costs $250 or roughly $450 to $1,060.

Google Gemini Enterprise. Not computable from published figures. The honest line for a budget is "ask sales for the per-seat price and the consumption overage, in writing, with the definition of a consumption unit."

The ranking on this one workload is tokens-plus-runtime (hundreds of dollars), then Microsoft credits (low thousands), then Salesforce credits (around ten thousand), then Salesforce conversations (tens of thousands). That is not a ranking of value; Agentforce comes attached to the CRM, Copilot Studio to the tenant, and the two API-first options come attached to nothing, which is the point of them and the cost of them.

What the numbers leave out

Who builds the agent. The two cheapest options on the table assume engineers. Copilot Studio and Agentforce assume admins. Price the people.

What happens when you run over. Microsoft's behaviour is documented: disabled at 125%, with an admin email. Salesforce's PayGo keeps running and bills you. Anthropic pauses a session at its budget with a `budget_reached` stop reason. OpenAI applies standard rate limits. Only one of these is a customer-facing outage, and it is the one most likely to be bought on a prepaid pack.

Governance obligations that now have dates. EU AI Act Article 50 transparency duties for systems that interact with people applied on 2 August 2026; the Digital Omnibus, published in the Official Journal on 29 July 2026, moved Annex III high-risk obligations to 2 December 2027. A customer-facing agent must disclose that it is an AI where the EU rules reach it, regardless of which platform meters it. The AI agent inventory and governance programme is where that lands.

What breaks in month three

The pack was sized on the pilot. The pilot ran 300 conversations a day; production runs 900; the prepaid capacity was never resized. On Microsoft that is a disabled agent. On Salesforce Pre-Purchase it is an unplanned true-up.

Nobody can reconcile the invoice to the telemetry. The vendor's unit (credit, action, outcome) is not a unit your observability stack emits. Build the mapping in month one: which of your logged events is a billable action, and what is the implied unit price. The agent evaluation harness should carry a cost-per-conversation assertion alongside the quality ones.

The model under the platform changes price. Anthropic launched Opus 5.5 on 22 September at $4 / $20, a tier below Opus 5's unchanged $5 / $25; Google's Argon has an unspecified introductory period. A buyer who pinned Opus 5 keeps paying $5 / $25 until they migrate; the saving comes from moving models, not from a price cut. On token-metered platforms a new, cheaper model shows up in your bill the day you switch; on credit-metered ones you see it only if the vendor re-rates the credit.

Next step

Run the same fifty tasks on two platforms for a week before you sign anything; the agent evaluation and regression harness is the procedure, and the enterprise voice agent contact centre workflow and multi-agent support deflection are the two shapes this agent usually takes. The free Enterprise AI Agent Platform Buyer's Guide has the back-office worked bill, the negotiation checklist and the scoring template; the AI governance and risk framework covers the Article 50 and Annex III timeline. If you want this arithmetic run on your own volumes, book an enterprise audit. Head-to-heads: Copilot Studio vs. Agentforce, Claude API vs. Copilot Studio. Tool pages: Agentforce, Copilot Studio, Vertex AI (the closest Google tool page we have; Gemini Enterprise has no page yet), OpenAI AgentKit, Claude API. Everything else for this audience is at the enterprise hub.