Tokens are now metered. How should agencies respond?
AI providers are moving from flat fees to pay-per-token. So what does AI really cost agencies, is it worth it, and should clients be paying for it?
Tokens are now metered. How should agencies respond?
AI providers are moving from flat fees to pay-per-token. So what does AI really cost agencies, is it worth it, and should clients be paying for it?
By Conor Nichols, Special Reporter
For the past few years, in some ways AI has been the cheapest colleague in the building. Agencies have bought licences, absorbed the running costs and in plenty of cases, used their AI offerings as a reason for clients to pick them.
Unfortunately, that arrangement is ending. AI providers are shifting from flat subscriptions to metered pricing, charging for every token a model reads and writes. GitHub, Microsoft’s coding assistant business, moved all Copilot plans to usage-based billing on 1 June 2026. The CEO of marketing industry consultancy Ebiquity, Ruben Schreurs, also told Digiday that agencies “can’t afford to continue to subsidise all those costs.”
This leaves agencies with some awkward questions, several of which have existed for years but are being forced to the surface…
Do they actually know what their AI costs? Is that spend earning its keep, or just making the same work faster? When AI goes into client work, should the agency absorb the cost or pass it on? Are the people using these tools every day aware that the meter is running? And if prices rise after agencies have cut fees on the promise of AI efficiency, then what?
We put these questions to Pete Robins, founder of IPA Accelerator agency Project5, Virginie Goupilleau, founder of AI consultancy BerylliumIV, and Karla Smith, Honorary Treasurer of the IPA and former CFO of Ogilvy UK.
A quick explainer on tokens and their pricing
A token is a fragment of text, roughly three quarters of a word, and AI providers are charging per million tokens, counting everything you send in and everything that comes back. Output usually costs four to five times as much as input.
Some of the meter is also invisible. Reasoning models “think” before they answer, and that thinking is billed too. A single request can burn tens of thousands of tokens before you see a word.
Image and video tools tend to use their own units, charging per image or per second. Google’s Veo 3.1, for example, bills per second of video. The principle is the same… the more you generate, the more you pay.
The confusing part of this is that token prices are falling and fast. A price war broke out over the summer, with OpenAI launching models at half the per-token cost of their model predecessors.
Yet bills are on the rise, for three reasons:
- Agents are hungrier. An agentic query can use five to 30 times the tokens of a chatbot, according to Gartner analyst Will Sommer.
- The newest models cost more. Some flagship models launched this year at twice the price of the ones they replaced.
- Terms change quickly. GitHub nearly quadrupled the cost of using its top-tier models within three months.
This free-sample-then-charge model is “the drug dealer’s marketing strategy”. So says author and management thinker Margaret Heffernan who also points to the squeeze agencies will recognise. Clients are already arguing that AI should make work cheaper. “Lower your prices, deploy AI - and then what’s the plan when prices rise?” she asks. “Hope is not a strategy.”
Back to the agency questions.
Do you actually know what you’re spending?
Virginie Goupilleau
Founder
AI consultancy BerylliumIV
Karla Smith
Honorary Treasurer of the IPA and former CFO of Ogilvy UK
Virginie Goupilleau, founder of AI consultancy BerylliumIV, says most do not. “The holding companies are way ahead on their AI journey and have rebuilt their operating models around it,” she says.
"The challengers and the longer tail are still building and piloting, and therefore still handle token spend like a production cost, either absorbing it or passing it through on a hunch. Our maturity assessments run this year confirm few of them measure usage with any precision."
It is not just agencies. In a KPMG survey reported by the Wall Street Journal, only 26% of companies said they have a comprehensive view of their AI costs, while 22% have no visibility at all or only find out what they owe once the bill arrives. KPMG’s Steve Chase described companies that exhausted their entire annual token and cloud computing budgets within months, and one client whose token usage jumped sixfold.
Karla Smith, Honorary Treasurer of the IPA and former CFO of Ogilvy UK, says agencies need to pause and look inward. “As we move forwards with embedding AI into our standard ways of working across creative, production and media agencies we need to take a step back and review the impact this is having on our own agency profitability,” she says. “Depending on your agency size, your choice of platform and usage model will vary, but the challenge of how to manage the investment is the same for everyone.”
Is it earning its keep?
Pete Robins
Founder
IPA Accelerator agency Project5
Pete Robins, founder of IPA Accelerator agency Project5, has taken a deliberately practical approach. “We’ve split AI into three obvious parts: stuff we’ve built, stuff we’ve paid to be built, and stuff we rent,” he says.
Project5 recently added an AI cost line to its P&L. “We estimate the value of the AI we use versus the people cost of each planner it helps enhance.”
"If the cost is less than the incremental value, great. If not, not so great."
Robins is also unsentimental about what off-the-shelf AI does for his discipline. “Despite some of the bluster, headline PR and notable wider industry successes, AI is only just beginning to help the media planning part of our industry,” he says.
"Generic AI has so far been just a bit ‘meh’ for great media planning, regardless of the quality of the prompt. LLMs live in the generic, or use industry practices which are now mostly out of date. Functional but no flair."
Ultimately, Project5 measures AI against what matters… whether it improves the planning, and in turn media effectiveness. “Part maths, part opinion,” Robins says. “I’ve been media planning for 33 years, so our judgement is something I’m happy to back. And I will soon build an agent to help me do that based on our rules.”
That judgement gets more valuable as the meter speeds up. “As AI moves towards usage-based pricing, the costs will go up, so that judgement becomes increasingly important if we’re going to protect or expand our margins.”
Is it a cost or a product?
Though it may seem obvious, the time AI saves has to go somewhere. Goupilleau points to BCG research finding 66% of frontline AI users get limited or no guidance on what to do with the time saved, and more than half do not redirect it into strategic work. “So where did it go? Into more of the same, or up the value chain: new products, closer strategic relationships with brands, new revenue streams?”
That, she argues, makes AI cost a strategic choice rather than a line on the IT bill. “If AI is optimising your own delivery, treat it as an operating cost: budget it, optimise it, manage the risk. If you’ve layered in value and IP the client uses, and you packaged it in your product offering, the client should pay for both the work and the usage behind it.”
"Either way, measure usage accurately and cost it transparently in client engagements.
Robins has already made that call. “What we build becomes our IP. And if it genuinely improves the planning and creates a competitive advantage for our clients, we fully intend to charge appropriately for it.”
Production companies have already started formalising this, with AI studio Silverside assigning clients a set number of tokens based on an initial cost estimate. If their needs exceed that amount, they top up with another tranche, according to Digiday.
Does the person at the keyboard know the meter is running?
This is Smith’s central point, and possibly the least discussed at the moment. “Key to this is not the client negotiation or the cost of the system you chose to implement, but your employees’ understanding of your strategy and how you want them to engage with the new technology that is at their fingertips,” she says.
"Do your staff know the cost of the platform you have chosen or how the token system works?"
“Have you shared your strategy with them? Did you talk them through why you have chosen the route you are following and why other routes didn’t make the cut? Have you asked for their views and opinions and listened to their answers? We are all still learning and good ideas are going to come from all levels in your teams.”
Smith calls for transparency with staff about the cost impacts. “Making sure everyone is on the journey together and you share the financial impact of the AI usage decisions you want your teams to make will help them understand that their daily choices make a difference to the financial outcomes of your agency. Break it down and tell the story of how each decision impacts the bottom line. The more information you arm your people with the more they can help impact the agency overall and working together will improve your bottom line.” "
How locked in are you?
On this, opinion is split.
In our September issue, Mediaplus UK’s Liz Duff warned that agencies risk becoming price-takers again, just as they did with the big platforms, by concentrating their capability and costs around a handful of providers whose pricing they do not control. Tools can also simply disappear, as happened when OpenAI shut down its Sora 2 video API on 24 September 2026, leaving anyone who built on it to start again.
Even locking in a good deal carries risk, with some holding companies buying tokens in bulk. Schreurs has warned that if agencies commit to large volumes at fixed rates and then get better at prompting or orchestration, they could need far fewer tokens than they have paid for. “They’re stuck with this massive amount of tokens,” he told Digiday.
Smith is more relaxed. She recognises that changing supplier pricing and the uncertainty over future price hikes is a worry, but stresses it is one shared by every agency and client. “The number of providers in the market means this will even out, and with a strong internal strategy and plan around how you want to use AI in your agency, moving from one provider to another is not as hard as you might fear.”
For Robins, part of the answer is owning more of what you use. That is where his split between AI you build, AI you pay to have built and AI you rent comes in. It also plays differently depending on agency size. “The bigger agencies will be better at some aspects, smaller indies like us at others. Our advantage is we don’t need industrial-scale enterprise AI that takes time to sort and roll out. We need smaller, more agile, built-to-our-spec stuff. Our disadvantage is we don’t have £300m to invest or a smart, dedicated AI team.”
The bill was always coming
This is not an argument against using AI but an argument against using it without knowing what it costs.
For Smith, the bigger threat to margins is not the AI providers at all.
"AI or not, the biggest challenge to agency margins is internal inefficiency," she says.
“Where the strategy agreed by the board is not understood or implemented by the workers on the frontline, this has a big impact on the bottom line. This is the same with how we will implement AI usage.”
Heffernan argues that in a period this uncertain, the goal should be robustness rather than efficiency. For agencies, that means knowing what AI costs, what it earns, and what you would do if the price doubled tomorrow.
Published by the IPA, the IPA AI Magazine helps member agencies make sense of the fast-changing AI landscape, with practical insights, expert perspectives and real-world applications of generative AI.
Managing Editor: Conor Nichols
Design: Jenna Betts
© 2026 The Institute of Practitioners in Advertising. All rights reserved. No part of this site may be reproduced without our permission.
