The subsidised phase of AI is ending. The flat monthly fee that let staff run unlimited queries through an AI assistant was never going to last; it was a way to win customers quickly, and that phase is now over. Usage-based billing is arriving across the major platforms, and advice firms that have quietly built AI tools into daily work will feel it at the next invoice if they don’t look now.

What is changing

AI providers are moving from flat-rate subscriptions, where light users effectively subsidised heavy ones, to billing per million tokens used. [1] Google cut its Ultra tier from $250 to $200 a month in May 2026 but added usage-based billing for heavy use. [1] GitHub Copilot made a similar change, and developers who had budgeted on the old model were unhappy about it. [2]

The cause is compute rather than greed. Monthly token volumes have grown nearly sevenfold, from roughly 480 trillion to 3.2 quadrillion. [1] Running frontier models at that scale can’t be absorbed into a flat fee forever, and what’s giving way is the idea that AI is a fixed overhead.

For an advice firm that has two consequences. AI costs become variable, and possibly hard to read, unless someone is watching usage. The tools staff use most (drafting assistants, meeting summarisers, research tools) may use far more tokens than anyone has assumed.

What a token costs, and where the cost lands

A token is roughly three-quarters of a word. Running a long client document through a frontier model like GPT-4o or Claude Opus costs around $5 per million input tokens and $25 per million output tokens. [3] That sounds trivial until you consider that filling a million-token context window, useful for analysing a whole client portfolio or a long regulatory document, costs dollars per request, [4] and a firm may be doing dozens a day with nobody adding up the total.

Smaller models look very different. Some are priced at $0.50 per million input tokens and $2.50 output. [3] Vendors claim some of these come close to frontier performance at a fraction of the cost, though that’s worth testing on your own tasks before relying on it. None of that means sensitive work should go to the cheapest model. It means that using the most powerful, most expensive model by default for everything, routine tasks included, is a budget decision worth making on evidence rather than habit.

Usage billing is also an audit trail

Usage-based billing raises governance questions. Which tools process which data? Who authorised it? Is anyone monitoring the volume going through each system? A Senior Manager under SM&CR could reasonably be asked those questions if something goes wrong.

It also helps answer them. A flat fee tells you nothing about what’s being processed. A per-token bill, read carefully, is a rough record of AI activity across the firm, which is exactly the map most firms are now being asked to build. Every AI product I run has a hard spending cap, and the bill it produces is one of the first places I look when I want to know what a system has been doing.

Where firms are exposed

The exposure falls in predictable places. Tools staff found and expensed themselves rarely appear on any list, so nobody knows whether they’re flat-rate or about to switch. Many firms couldn’t pull thirty days of usage data if asked, which is a governance gap before it’s a cost one. Routing routine drafting through a frontier model is a cost choice, and sometimes a data-minimisation question too, because sending more data to a more powerful external model isn’t always right for a regulated firm.

Vendor viability belongs in due diligence as well. Several application-layer AI vendors are under margin pressure as compute costs rise, and a tool your operations team relies on today might move upmarket, put prices up sharply or close.

Owning the layer above the model

Chamath Palihapitiya put it plainly in a May 2026 discussion: “controlling the tokens is controlling the leverage.” [5] For an advice firm that means the part worth owning is the layer that decides what the model can see and produce, not the model itself. Build workflows directly on one vendor’s product and switching when prices change is painful. Build around your own data and process rules first, and the model underneath becomes a procurement choice.

The same logic favours grounded AI, where the model answers only from the firm’s own knowledge base. It keeps errors down, keeps data handling auditable, and makes the firm’s knowledge the asset rather than the vendor’s model.

My view is that usage-based billing will be the norm for every serious AI tool within eighteen months, and that the firms which come out ahead won’t be the ones that found the cheapest model. They’ll be the ones that can read their AI bill line by line and say what each line was for.

This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.

Sources

  • [1] Google, Google One AI Premium pricing update, May 2026. Supports the price reduction and usage-based billing shift for heavy users. Vendor-sourced.
  • [2] GitHub, GitHub Copilot billing model announcement, May 2026. Supports the transition to token-based billing and developer response. Vendor-sourced.
  • [3] Anthropic / OpenAI, Published API pricing pages, May 2026. Supports the $0.50/$2.50 vs $5/$25 per million token figures. Vendor-sourced.
  • [4] Various, AI provider context window documentation, 2026. Supports the dollars-per-request cost of million-token windows.
  • [5] Chamath Palihapitiya, All-In Podcast, May 2026. Attribution of the “controlling the tokens” framing.