Don't trust the PowerPoint or proposal email. Underneath the familiar wrapper, three genuinely different commercial models are competing for your budget, and confusing them is how some organisations and their procurement teams end up with a five-figure surprise on an invoice that was supposed to be predictable.
This article is not meant to question the significant value in adopting AI in many organisations, but with tighter budgets and prioritisation of spend on real value vs experimentation, being clear what the cost levers are is important to as part of a fully aligned adoption and change strategy.
Here's the landscape as it actually stands in mid-2026.
1. Per-seat SaaS: pay for access, not usage
This is the model everyone recognises: a flat monthly fee per user, largely decoupled from how much that user actually consumes.
ChatGPT Business runs $20 per seat per month on annual billing ($25 monthly, following OpenAI's April 2026 cut from $25/$30), with a two-seat minimum - you're buying a license, not metered compute.
GitHub Copilot used to sit in similar territory - and no longer does: since 1 June 2026 every GH Copilot plan bills on usage-based GitHub AI Credits, with GH Copilot Business at $19 per seat including $19 of monthly credits (1 credit = $0.01, consumed at each model's token rates). Code completions stay unmetered; agent runs and premium-model calls draw down the pool. It's a useful warning shot - one of the most-cited per-seat AI products in the market quietly became a hybrid.
The appeal is obvious: budgeting is trivial, procurement can forecast and negotiate in a P*V approach, and finance doesn't need a FinOps function to sanity-check the bill. The catch is that "unlimited" rarely means unlimited - most per-seat plans carry soft caps, rate limits, different usage depending on LLM model, or a separate metered layer for the heaviest features, which is where model two and three start creeping in even on products sold as pure SaaS.
2. Consumption-based: pay for what you burn
At the other end sits raw token or credit metering - the model underneath every foundation model API. You pay per million tokens processed, with output typically priced several times higher than input (a 5x ratio is common across current-generation models). There's no seat fee, no user count, no floor. A single power user can outspend fifty light users combined.
This is the model procurement teams understand least well and fear most, because it inverts the normal budgeting logic: cost tracks behaviour, not headcount, and behaviour is the one variable nobody controls precisely. It's also, not coincidentally, the model every hybrid structure below is quietly built on top of.
Examples are Microsoft Foundry, AWS Bedrock, Google Vertex from the main hyperscaler platforms - and the negotiation is commonly based on committed volumes / spend over time and ideally paying for capacity in advance rather than unit rates.
3. Hybrid: the model doing most of the damage to buyer clarity
This is where most current AI commercial constructs actually live, and it comes in two distinct flavours that get conflated constantly.
Flavour A - seat with included capacity, then overage. You pay a subscription that bundles a dollar-denominated allowance of usage, and once you exhaust it, you either upgrade or start paying metered rates on top.
Cursor is the cleanest example: a Teams Standard seat is $40 per user per month ($32 on annual billing), and since June 2026 each seat carries two separate pools of included usage - one generous pool for Cursor's own models (Composer, Auto), and a much smaller one for third-party models like Claude and GPT, billed at public list API rates plus a $0.25-per-million-token Cursor rate. It's the second pool that empties. Exhaust it and on-demand usage bills in arrears, or you move the user to a Premium seat at $120 ($96 annual) for 5x the allowance.
Microsoft 365 Copilot follows a related pattern with an extra wrinkle: the $30 per-seat licence is an add-on, not a product — it requires a qualifying Microsoft 365 base plan underneath, so the true all-in figure is closer to $66–$90 on E3/E5. That seat gets you the Graph-grounded core experience, but agent and advanced-feature work draws down a separate pool of Copilot Credits, priced at $0.01 each pay-as-you-go or $200 per 25,000-credit pack per month. Under 300 seats there's a cheaper lane again: Copilot Business at $21 per seat (promotional $18), or the bundled "with Copilot" SKUs at $23.50–$32 all in.
Flavour B - seat as a gate, not an allowance. Here the subscription doesn't include usage at all; it simply unlocks the right to be billed for usage.
Claude Enterprise is the sharpest current example: the seat fee is $20 per user per month billed annually — 20 seats minimum self-serve, 50 sales-assisted - and that fee includes zero usage - every token your organisation consumes across chat, Claude Code, and Claude Cowork is billed separately at standard API rates, with spend caps as the only real control lever. Claude Cowork specifically requires an underlying paid seat (Pro, Team, or Enterprise) before it's available at all, and because Cowork's multi-step autonomous tasks are compute-intensive, they draw down usage allowances - or run up metered costs - far faster than an ordinary chat session does. You're not paying for capacity; you're paying for admission, and the meter starts the moment you walk in.
The distinction between A and B matters enormously for cost modelling. Flavour A gives you a buffer before the metered exposure starts. Flavour B gives you none - the seat fee is closer to a platform access fee than a pricing plan, and the real number only appears once you've modelled actual consumption.
Why this matters more than the pricing page suggests
The vendors have strong incentives to let these models blur together, because "per user per month" sounds safe and "billed at API rates" sounds like a warning label - even when, structurally, the second is what you're actually buying. A seat number in a slide deck tells you almost nothing about total cost until you know which of these three (or more, counting the sub-flavours) constructs sits behind it.
For anyone advising on or procuring these deals, the practical implication is straightforward: stop comparing seat prices and start comparing what the seat fee actually includes. Ask vendors directly - does this fee include usage, and if so how much, denominated in what unit, resetting on what cycle? What happens at the boundary - hard stop, degraded service, or automatic overage billing? And critically, whose usage pattern was the pricing designed around, because a seat priced for occasional chat behaves very differently once agentic, multi-step workflows like Cowork or Copilot Studio agents enter the picture - those tasks consume tokens at a materially higher rate than a single question-and-answer exchange, and that's exactly where hybrid models turn from convenient to expensive.
The commercial model is no longer a footnote to the AI buying decision. Increasingly, it is the decision. Clients need to understand the shared responsibility of who controls the cost
The client (personas / profiles and their tasks, token usage, model choice, etc) or
The vendor (what volume commitment of users and tokens are you prepared to pre-commit to.
Sources (verified September 2026)
OpenAI — ChatGPT Business pricing ($20/seat annual, 2-seat minimum; April 2026 cut)
GitHub — "GitHub Copilot is moving to usage-based billing" (AI Credits from 1 June 2026; Business $19/seat incl. $19 credits)
Cursor — Teams pricing docs (Standard $40 / Premium $120, split usage pools, Cursor Token Rate)
Cursor — "Improvements to Teams Pricing", June 2026 (two usage pools, annual $32/$96)
Microsoft — Microsoft 365 Copilot pricing (add-on to a qualifying base plan; Copilot Business SMB tier)
Microsoft — Copilot Studio / Copilot Credits pricing ($0.01 per credit; $200 per 25,000-credit pack)
Anthropic — Claude pricing (Enterprise $20/seat annual, usage billed at API rates; Cowork included on paid plans)
Anthropic — API pricing (input vs output token rates, ~5x output multiple