The meter under the seat

AI is being repriced, and in 2026 the repricing did not arrive through the price. It arrived through the quantity included underneath the seat. I modelled what that does to one 400-seat contract from both sides of the invoice: the customer cut a quarter of its seats and the bill still rose 28 per cent, while the seller's margin fell nearly thirteen points on the same event. Here is what I am seeing as at September 2026, where I judge the market sits on the adoption curve, what a finance leader should do about it, and why I read it through two precedents I know from the inside: O2's handset subsidies, which I helped make measurable in the early 2000s, and the move to the cloud. My read is that most of AI is early in its growth phase, that the light user and now the advertiser are paying for the heavy one, and that as the meter arrives there will be winners and losers on both sides of the invoice, as there were in mobile and in the cloud.

A note on scope. The AI half is a practitioner's read of published pricing pages, analyst data and my own invoices, as at September 2026, with judgements flagged as judgements; I am a paying customer of the products named and have no commercial relationship with any vendor. No vendor publishes what it costs to serve a subscription, so nothing here asserts one, and every cost figure in the worked model is invented. The mobile half is first-hand, from inside O2's consumer division, anchored to the public record where it matters; the disciplines are far older than my involvement in them.

In this piece:

1. What I am seeing

2. Where the market is on the curve

3. What the meter actually charges for

4. One account, two contract designs, two years

5. Two precedents: mobile, and the cloud

6. Winners and losers, now and after the meter

7. The seller's model

8. The buyer's mirror

9. What to do now, on either side of the invoice

10. In brief, and the model to download

The one-page model behind the worked example is downloadable here (Excel, inputs unlocked). Two earlier pieces carry the ground this one stands on: What AI actually changes in FP&A and The SaaS metric set is the patient funnel wearing different clothes.

What I am seeing

On 1 September 2026 the price of a GitHub Copilot seat did not change. The compute included inside it fell by more than a third on Business seats and by 44 per cent on Enterprise. That is the story of AI repricing in 2026 in one line. It was not a price rise, so it will not appear in any price variance, and it landed on the desk that manages delivery, not the one that manages the budget. GitHub is the running example in what follows for a reason: it is the clearest published case I have found this year, with the credits inside the seat, the overage rate, the pooling and the billing timing all on the record, and it is sold into the kind of organisation most of my readers run.

Three other things happened in the same window, and together they describe a market changing shape. Google launched Gemini 3.8 Flash on 2 September at $0.75 and $3.75 per million input and output tokens as introductory pricing through 31 December, rising to $1.50 and $7.50 from 1 January 2027: a 100 per cent unit-price increase already on the 2027 calendar. OpenAI cut its GPT-5.6 Sol model to $4 and $20 per million on 21 August, a promotional rate it has committed to hold at least until 21 November, then launched GPT-6 Astra two weeks later at $10 and $50, the newest tier priced above the last. And Ramp's September index of roughly seventy thousand US businesses shows the biggest spenders cutting AI spend per employee by nearly a tenth in August, to $7,205, with the effective price they pay per million tokens down 41 per cent since March, to $0.68, as the vendors cut prices and the buyers routed routine work away from frontier models. A fourth is already dated: from September, as GitHub reopened Copilot Business and Enterprise sign-ups for customers paying by card or PayPal, every new seat has had to be paid for before it is switched on, and from the billing cycle starting 1 October it charges those customers for every assigned seat in advance, prices unchanged. Cash timing is the third lever after quantity and mix, and it reaches the card-paying installed base on 1 October; check which of your own entities pays which way.

Read those together and my judgement is this: the flat-rate era of AI is ending, but not the way most budgets assume. At the vendors I buy from and the ones I have watched most closely this year, the rate has been the smallest mover; other suppliers run different structures, and the pattern is an observation, not a law. The included allowance, the model mix and the volume of work being pushed through the meter are the big ones, and only the allowance is entirely out of the buyer's hands; the mix and the volume are the buyer's to manage, where the product and the architecture expose them. List rates did fall this year, Sol's input rate by a fifth, Sonnet 5's scheduled increase cancelled, cached input by three quarters, and the landed bills I see still rose, because the allowance and the volume moved further the other way. A 2027 plan built on today's frontier rate is built on a promotion. A 2027 plan built on seats is measuring the wrong unit.

Where the market is on the curve

Subscription businesses are routinely misread at exactly this point, and I have watched it happen once before. In early adoption most subscribers use a fraction of what they pay for, a few heavy users cost far more than any meter would have charged them, and the two net off well enough that a flat rate looks sustainable and the land grab makes sense. In the growth phase the heavy users multiply, because the product has stopped being a novelty and become the workflow; the netting stops working, and the pricing has to change shape. That is the phase most of AI entered in 2026, with the embedded products already behaving like mature ones, and the allowance is how the shape changed. In the mature phase the product is simply embedded, a cost of doing business rather than a decision anyone revisits, and the buyer's leverage moves from the price to the architecture: whether you can switch what sits behind the interface.

Look at how AI is still sold today and the early-adoption shape is visible: a flat monthly subscription across users whose usage differs by an order of magnitude, so that wherever cost follows usage the light user is carrying the heavy one inside the same invoice. Hardware trained us to accept that shape. The iPhone 18 Pro went on sale on 18 September 2026 with a variable-aperture camera, the feature Samsung dropped in 2020 after two generations; my judgement is that a minority of buyers will ever reach for the manual control it adds, everyone pays for it, and the phone will sell. A flat AI seat is that camera with one difference: the light user is paying for someone else's use, not for a capability of their own. Whatever gap exists between what the heaviest users consume and what they pay, the provider carries it; the size of that gap at any named vendor is not public, the worked example below shows only how such a gap arises. What is on the record is sellers saying it about themselves. In January 2025 OpenAI's chief executive wrote that the company was "currently losing money on openai pro subscriptions" because "people use it much more than we expected". In August 2026 Canva, as the Australian Financial Review and Startup Daily reported, cut its revenue growth guidance for the year from 30 per cent to 20 per cent, its chief executive saying that the average cost of serving an AI task had been too high, that the company had leant too heavily on frontier models, and that its pricing, consumption model and usage controls had not caught up with demand. By 19 August its co-founder was telling Forbes the company had cut those serving costs by about 90 per cent, largely by building its own models, which is the other way out of a subsidy and one the mature phase will see more of. Those are the operators of the subsidy describing it from the middle.

What I judge pushes AI through the growth phase faster than mobile went is an asymmetry every experienced finance reader will already be holding. A handset subsidy was bounded: the operator knew what the phone cost on the day it handed it over, the number never grew, and the recovery ran on a fixed contract. A compute subsidy is recurring, variable and not bounded by the contract. A single power user, or one agent left running overnight, can multiply the cost of serving a flat-rate seat until the seller reaches for a limit, a throttle or a cap, which is precisely what the sellers reached for in 2026. The handset subsidy was a capital cost per acquisition; the compute subsidy is an operating cost per use, on a meter the customer holds. And the driver has a name. Agentic workloads, where the software runs a task end to end rather than answering a prompt, consume many times what a chat exchange does, because every step, retry and tool call is billed. Gartner's August 2026 forecast is that inference cost per agentic workflow will rise more than fivefold through 2028, an "inference paradox" in which the unit economics improve while the total bill escalates. Uber found the enterprise version this year, as Forbes and TechCrunch reported: coding-tool adoption climbed from 32 per cent to 84 per cent of an engineering organisation of roughly five thousand between February and March, helped along by an internal leaderboard, the company had spent its entire 2026 budget for those tools in four months, and by June it had capped spend at $1,500 per employee per month per tool.

Which is why the allowance exists at all. A seller facing an unbounded serving cost cannot price a flat seat and hope. It has to put a quantity inside the seat and decide what happens above it. That quantity is now the most important number on the pricing page, and in my experience most buyers have not read it.

I say this from three seats. The first is O2's, which comes later in the piece. The second is the owner's: for eleven years I have paid customer-acquisition cost out of my own pocket, with no balance sheet to hide it in, which teaches a truth that scale can obscure. A subsidy is a bet on lifetime value, and when the capital is yours you cannot place that bet on faith. The third is the buyer's: I run AI inside a live profit and loss today, in daily use across analysis, forecasting and modelling, on a bill that is small, fixed and, for what I put through it, generous. To a finance eye that generosity is not a gift. It is a tell, and reading it as one is my judgement rather than anything a vendor has published.

There is a second way to fund the gap between what the heaviest users cost and what the lightest pay, and it arrived this year: a second payer. Mobile never had one; the handset was recovered from the subscriber through the tariff or not at all. In AI the free tier has found an advertiser. OpenAI began showing ads to users on ChatGPT's Free and Go plans in the United States on 9 February 2026, brought them to the UK in June, its first European market, and to 31 further European markets in late August; by 31 August its head of ad solutions was quoting a $1bn annualised run rate reached in under two hundred days. Google already owned the platform: ads sit inside AI Overviews in twelve countries, the UK not among them on Google's own list as I write (ads above and below an Overview run everywhere), Gemini-built formats for AI Mode were announced in May, and Google had not announced advertising in the Gemini app as I write, Alphabet having told analysts in April that a format which works in AI Mode could transfer to it. Anthropic went the other way, pledging on 4 February 2026 that Claude will remain ad-free, paid for by subscriptions and enterprise contracts.

A second payer extends the curve. An advertiser funding the light users lets a vendor hold a free or near-free tier open long after the seat economics alone would have forced a change of shape, and the longer it stays open the deeper the product embeds in the daily workflow, which is the switching cost any later price rise or rival's cheaper seat has to beat. It also splits the customer base: advertising pays for the consumer tier, nobody funds an enterprise seat with a sponsored link, so the paid tiers carry the meter in full and the allowance matters more there, not less. And it makes vendor choice a question of who else is paying for your seat. Three suppliers now hold three positions on that, two declared and one left open, and my judgement is that the position says as much about where a vendor's prices are heading as the price list does.

What the meter actually charges for

Five charging units are live as at September 2026, and each allocates risk differently. A seat charges for access and leaves the seller holding the serving-cost risk. Pure usage meters every unit and leaves the buyer holding the demand risk. A hybrid, a seat with an allowance inside it and a rate above it, splits the two. Committed-spend contracts prepay a pool of usage at a discount and move the demand risk onto the buyer's balance sheet. Outcome pricing charges for the result, which sounds like the buyer's dream until you ask who pays for the attempts that failed. The headline structure is not the decision. The allowance design underneath it is: per seat or pooled, rollover or expiry, and the overage rate against the cost to serve. Those three settings move margin more than the headline price does, and the worked example below shows by how much.

GitHub's is the clearest published illustration. On 1 June 2026 it moved every Copilot plan onto AI credits at a cent each, metered on tokens at published rates, with pay-as-you-go above the allowance, and the credits are pooled across the billing entity rather than ring-fenced per seat. Read the plans page and a pattern appears in the base allowances: they equal the seat price to the dollar, $19 and 1,900 credits on Business, $39 and 3,900 on Enterprise, with the individual plans adding a flexible top-up on top. The base seat, in my reading, is a prepaid compute account with a seat label on it. Then on 1 September the promotional allowances ended: Business seats went from 3,000 included credits to 1,900 and Enterprise from 7,000 to 3,900, with the seat prices unchanged at $19 and $39.

Across the market the hybrid is now the most common structure, on two populations that should be read separately. Among 230 B2B software companies polled in May 2026 it was the primary model at 37 per cent, up from 25 per cent a year earlier, a plurality; among AI sellers specifically, a Salesforce Ventures survey of more than 300 executives found over half on a flat base with consumption above a cap. Bain's August 2026 analysis of roughly 200 B2B software companies' published pricing found four in five vendors introducing AI pricing choosing capacity models, which I read as capacity preserving the economics of the seat, with about one in five AI-native companies still mostly per-seat. Outcome pricing, for all the conference talk, was about a tenth of the metering Bain counted; it has taken hold in customer support, where Intercom charges per outcome and Zendesk per automated resolution, and is still uncommon elsewhere in Bain's sample. Anthropic, meanwhile, cancelled a scheduled increase on its mid-tier model, cut the price of cached input by three quarters on its two newest ones, and on 14 September reset the weekly limits on my own coding plan: a permanent 25 per cent rise on the standard limit, and, measured against the summer's temporary boost, a 17 per cent cut, which the vendor itself said. Two baselines, one change, and the one you notice depends on which you were using. Vendors are discounting the cache and adjusting the quantity, not the sticker.

One shape I expected to be still ahead of us arrived first. AI folded invisibly into a product you already buy, its cost recovered somewhere you cannot see, is not the future. Microsoft moved Copilot into its consumer Microsoft 365 plans in late 2024 and early 2025 with price rises attached, Google folded Gemini into its Workspace Business and Enterprise plans at the start of 2025 and raised the base price, and on 1 July 2026 Microsoft's commercial prices rose at each customer's next renewal, Office 365 E3 from $23 to $26, up 13 per cent, with AI capabilities in the packaging. A CFO still treating embedded AI as a future problem has already paid for it, on a line item that is not in the AI budget. The mature phase arrived early, inside products that were already mature, and it arrived with a price rise.

One account, two contract designs, two years

Invented round numbers throughout. Northwind is not a real company, the figures are not anyone's data, and the cost to serve is an assumption, because no vendor publishes one.

The example in three numbers. The same 400 seats, the same usage and the same cost to serve earn the seller a 59.9 per cent gross margin if the allowance is granted per seat and 48.4 per cent if it is pooled across the account: one contract line, more than eleven points. In year two, when agent runs replace chat sessions, the customer cuts a quarter of its seats and the bill still rises 28 per cent, while the seller's margin falls from 59.9 to 47.0 per cent on the same event. And the acquisition payback that reads as 4.0 months on revenue is 6.6 months on gross profit. The working follows, step by step; the model behind it is downloadable at the end of the piece.

The set-up. Northwind sells 400 seats to one customer at $40 a seat a month, including 1,000 credits, with overage at three cents a credit. Its cost to serve is two cents a credit; both rates are assumptions, chosen so that overage earns a third of its price as margin rather than the near-total margin of the software era. The consumption looks like this:

Ten per cent of the seats drive 53 per cent of the compute. The mean of 1,140 sits above the 1,000 allowance because of that tail, and an allowance sized on it would be wrong for both ends of the distribution.

Design A grants the allowance per seat. The light cohort pays $9,600 and costs $1,440 to serve, an 85 per cent margin. The regular cohort pays $5,520 including overage and costs $2,880, a 48 per cent margin. The heavy cohort pays $7,600 and costs $4,800, a 37 per cent margin. In total: revenue $22,720, cost $9,120, gross profit $13,600, gross margin 59.9 per cent.

Design B pools the same allowance across the account. That is 400,000 credits for the customer as a whole. Consumption is 456,000, so only 56,000 credits are billable. Revenue $17,680, cost $9,120, gross profit $8,560, gross margin 48.4 per cent.

Same customer, same usage, same cost to serve. Put the two totals side by side: revenue $22,720 under Design A against $17,680 under Design B, a difference of $5,040 a month and $60,480 a year, and because the cost to serve is the same $9,120 in both, the whole $5,040 is gross profit, $13,600 against $8,560, which is the gap between a 59.9 and a 48.4 per cent margin. One line in the contract is worth more than eleven points of gross margin. And it is not hypothetical: GitHub pools Copilot credits across the billing entity, which puts it on the Design B side of that line.

Where the subsidy sits. Blended cost to serve is $22.80 per seat against a $40 seat, a 43 per cent margin before a single credit of overage. But a heavy user costs $120 a month to serve against the same $40. Under Design B the heavy cohort's seat revenue is $1,600 against $4,800 of cost to serve, a contribution of minus 200 per cent on seat fees alone, funded by the light cohort's plus $8,160: the subsidy did not disappear when the seller added a meter. It moved inside the customer, which is early adoption in one table.

Year two, growth: the number most budgets leave out. Let regular volume grow 20 per cent and heavy volume 40 per cent in year two: $3,744 a month of expansion revenue, costing $2,496 to serve, an incremental gross margin of 33 per cent against something close to 100 per cent in the software era. That single line is one reason AI product margins are reported in the fifties against software's eighties, and why a payback struck on revenue is wrong by two fifths to just over a half: on the year-one run rate a $90,000 acquisition cost pays back in 4.0 months on revenue but 6.6 on gross profit under Design A, and in 5.1 against 10.5 under Design B. Same cost, same customer. The contract line moved it by nearly four months, the wrong denominator hid between two fifths and just over half of it, and a survival curve, which any real cohort has, would stretch both figures further.

Year two, the mirror: the growth phase in one table. Run year two with an agentic shift instead. The customer cuts a quarter of its seats, the hundred lightest, who were barely using it, leaving 140 light, 120 regular and 40 heavy. But agent runs replace chat sessions: regular consumption doubles to 2,400 credits a seat and heavy rises to 11,000, with the remaining light seats held at 300. Under Design A the bill is $5,600 plus $9,840 plus $13,600: $29,040 a month. Seats down 25 per cent. The bill up 28 per cent. $272,640 a year becomes $348,480, and cost per seat rises from $682 to $1,162. A budget built on seats misses all of it. And the seller does not celebrate either: cost to serve rises to $15,400, gross profit is $13,640, and gross margin falls from 59.9 to 47.0 per cent even though revenue rose 28 per cent. Decompose it and the mechanism is plain: losing the hundred lightest seats gave up $4,000 of revenue at an 85 per cent margin, minus $3,400 of gross profit, while the overage growth added $10,320 of revenue at 33 per cent, plus $3,440. Net movement in gross profit: forty dollars, on a 28 per cent rise in revenue. The buyer's bill shock and the seller's flat profit are the same event. Run the same year under Design B, the pooled design GitHub uses, and the shock is larger, not smaller: the pooled allowance shrinks with the seats it is attached to, so the bill rises 48 per cent, from $17,680 to $26,100 a month, while the seller's margin falls from 48.4 to 41.0 per cent.

The four buyer levers, sized on these numbers. Route routine work to a cheaper model, which is part of how Ramp's buyers found 41 per cent; move the allowance from per seat to pooled, worth 22 per cent of the bill; cap the tail; and negotiate carry-forward and a tiered overage into the order form.

Two precedents: mobile, and the cloud

The reason I read the AI market this way is that I watched the same shape from inside the last one. A mobile operator once handed a stranger a costly handset for nothing upfront and set out to earn it back over the life of the contract. It looked like madness; it was a disciplined piece of unit economics, and it only worked if finance could tell which subscribers were worth the subsidy. The OECD, looking back in 2013, said the arrangement was "commonly but inaccurately described as a handset subsidy" and was "rather a bundled sale": a lower price for the device in exchange for a binding contract, the difference recovered through the monthly fees. My own gloss is blunter, and the OECD's paper reaches for the same word: embedded consumer credit, with the interest hidden inside the bill. A Copilot seat whose base credits equal its price is the same trick with a different label.

At O2, across pre-pay customers and post-pay contracts, I built part of the finance side of that measurement. I worked with IT on the SQL that pulled per-line usage data, voice in and out, SMS, MMS, data and spend, adjusting the scripts myself when an extract came out in the wrong shape or fell over on the volume, and joined the result to the device, the tariff and the acquisition channel. Out of that came a customer net present value by tariff, by device and by sales channel, with churn computed by channel and by tariff as an input to it. A customer acquired through a given retailer, on a given device and tariff, carried a modelled value profile before they had made a second call. The commercial teams used it to redirect live marketing spend towards the profiles worth acquiring, and I partnered them to 33 per cent profitable sales growth over my time there: the growth was theirs to win, and the profitability view told them where to aim. Churn and average revenue per user were operating metrics in mobile before software borrowed the vocabulary, and what mobile then called subscriber acquisition cost is what everyone now calls the cost of acquiring a customer.

Mobile also shows how this tends to finish, because it is twenty years further through. Metering was never the new thing: UK contract tariffs of the period, the ones Oftel's 2001 competition review examined, already carried an inclusive-minute allowance with a per-minute rate behind it. What changed, slowly, was that the acquisition subsidy was prised out of the recurring bill. SIM-only gave the disciplined customer an opt-out. Data allowances put a meter on the thing people had actually started using. And only late in the cycle were the handset and the airtime formally priced apart: O2 Refresh, launched in April 2013, was, as the European Commission recorded it in the Three and O2 merger decision, "the first 24 month tariff to decouple the cost of the phone from the cost of mobile services". A decade after I left. Even then the customer had been trained: in July 2019 Ofcom found two million bundled customers sitting out of contract, around 1.4 million of them paying more than a comparable SIM-only deal would have cost, some GBP 182m a year between them, because a bundled bill has no handset instalment to stop paying, while more than a quarter of the two million were better off where they were. Split contracts were by then 15 per cent of the pay-monthly market, sold by five providers; EE, Three and Vodafone had none. Not finished. Just twenty years ahead.

That regulator's argument, that bundling a component into a subscription is what stops the customer seeing its price, is no longer a precedent for AI. It is a live case. The Competition and Markets Authority opened a consumer-protection investigation into Microsoft in late July 2026 over whether customers were clearly told about the lower-cost option when Copilot was added to Microsoft 365 consumer plans; Italy's competition authority opened its own inquiry in June, and Australia's consumer regulator has the same conduct in the Federal Court, over Microsoft's communications to about 2.7 million subscribers. None of the three has decided anything; they are investigations and allegations, not findings. But the question they raise, whether the price of a bundled component was made clear to the customer, is the one Ofcom examined in 2019.

Cloud computing ran the same curve fifteen years ago, and it is the nearer precedent for the mature phase. Servers you owned became capacity you rented by the hour, the first bills arrived as a shock, and the shock created a discipline called FinOps: usage tagged to owners, committed-use discounts traded against flexibility, egress charges read before the headline rate, and multi-cloud designs that exist because nobody wanted to depend on one provider's price list. Deloitte's Finance Trends report, published in September 2026 from a spring survey of 1,434 finance leaders at companies with revenue above a billion dollars, found 60 per cent expecting to need more sophisticated AI cost management through 2027, and a FinOps capability in place at 38 per cent of those already preparing for it, against 25 per cent of those maintaining current practices. The discipline the cloud forced on IT budgets has not reached most AI budgets yet. It will. Neither mobile nor the cloud rewarded the early bargain for long; both rewarded the side that could measure and could move.

One disanalogy worth keeping honest. Mobile's lock-in was contractual; consumer AI lock-in is habit on a rolling plan, and a November 2025 survey of two thousand paying US subscribers found them paying for four tools at almost $66 a month, which is a portfolio, not a commitment. Enterprise lock-in is a different animal, workflow integration, accumulated context, evaluation suites and agent configuration, and it hardens every quarter: the consumer negotiates by leaving, the enterprise has to negotiate with data and with an architecture that can switch.

Winners and losers, now and after the meter

Put the curve, the example and the two precedents together and the market sorts into winners and losers twice. What follows is my judgement, by role rather than by name.

Today, early in the growth phase, the winners are the heavy users on flat plans, whose consumption the flat rate was never sized for, and the vendors buying share with that subsidy, because a workflow, once embedded, is expensive to leave. The losers are quieter: the light users paying for capacity they never touch, which is the same invoice read from the other end, and the sellers' margins carrying the difference. The second payer changes who funds the consumer tier; it does not change the arithmetic on the enterprise one, and it creates a loser of its own, the paid consumer tier, now competing with a free tier that someone else is paying to improve.

As the meter arrives, the positions flip. On the buyer's side the winners are the finance teams that instrumented usage while it was cheap, know their own tail, and can route a task to a cheaper model or another provider on a configuration change; the losers are the budgets built on seats and today's rate, which meet the next allowance change as a delivery problem in week eight of a quarter. On the seller's side the winners have a second payer, their own models or deep enterprise integration, each of which funds the subsidy or makes it unnecessary; the losers anchored customers to a flat rate they cannot sustain and will spend years walking it back, as the mobile operators did, with a regulator at the end of it. Mobile took twenty years to finish that sorting. My judgement is that AI will do it in a fraction of the time, because a compute subsidy runs until the seller caps it, and the sellers reached for caps within a year.

Which side of the line you end up on is decided now, by what you build while someone else is still paying for the habit. That is what the two models that follow, and the mandate after them, are for.

The seller's model

If you are the CFO of the company doing the pricing, the model has five components, and they feed each other in this order. I have set terms from the seller's side as well as measured from the buyer's. At Chiron I renegotiated the distribution agreements across a seventeen-country network, modelling each partner's own profit and loss to find the share both sides would sign, because a term you cannot price for the other side is a term you will not get; and for eleven years I have set my own price points against a marketplace that moved the terms without asking. An allowance and an overage rate are the same negotiation in a different costume. Read the five as the order in which the numbers feed each other: the charging unit decides what the distribution does to margin, margin decides what retention is worth, and retention decides what an acquisition can pay back.

1. The charging unit, and the allowance design underneath it. Write down who carries which risk under each structure: demand, serving cost, adoption and, for outcome pricing, the definition of an outcome. Then decide the settings that actually move margin: per seat or pooled, rollover or expiry, and the overage rate against cost to serve.

2. The consumption distribution, never the average. Light, regular, heavy and automated cohorts, and the spread within them. In AI workloads the tail is the profit and loss, and an allowance sized on the mean sits above most users and below the ones who matter.

3. Gross margin after inference, including the incremental margin on expansion. Allocate variable cost to chargeable and non-chargeable events alike, and if you price by outcome, cost the attempts that failed, because outcome pricing transfers that cost to you rather than removing it. The benchmark is not the software margin investors grew up with: ICONIQ's 2026 survey of more than 300 AI-product executives put gross margin at 45 per cent in 2025 and a projected 53 per cent this year, a long way below the roughly 80 per cent that mature software companies have typically reported.

4. Cohort retention and net revenue retention, with the denominator stated. Usage-based companies posted a median of 108 per cent on 2025 actuals against 98 per cent for seat-based, and the ratio moves with contract design, as the worked example above shows: in its mirror year net revenue retention is 128 per cent on a 25 per cent seat loss, which is the contract design talking, not the customer relationship. I say this as current practice rather than anything I ran at O2, where the discipline was churn by channel and by tariff, one layer down.

5. Customer-acquisition payback struck on gross profit after serving cost, never on revenue, with cohort survival in the model; the survey median is 16 months, the top quartile six or fewer. The accounting sits underneath all five: fixed versus variable consideration, the breakage on unused credits, which is revenue with no serving cost once the contract and the revenue standard let you recognise it (in the worked example the light cohort leaves 168,000 credits unused under Design A, which is where its 85 per cent margin comes from), and the seller's own prepaid compute commitments, which are the closest true analogue of the handset subsidy in this whole story.

The buyer's mirror

Now the same model from the other side of the invoice. A finance leader budgeting AI spend for 2027 needs a decomposition, not a multiple:

Landed cost = seat fees + overage, where the overage is the sum, across every active user, of the usage above that user's allowance, priced at the rate of the model that served it; a pooled allowance nets across the account before anything is billable. Five levers move it: the rate by model, the volume per active user, the number of active seats, the model mix, and the allowance design.

What moved in 2026, in order of size: the allowance (GitHub cut what is included and left the price alone); the model mix (Ramp's buyers cut their effective token price by 41 per cent, partly by routing routine work to cheaper models and partly because the vendors cut prices); volume and adoption (the agentic driver, and Uber's four-month budget); and last the rate, which barely moved this year and has dates attached for next. Today's frontier rate is a promotion with an expiry date, which is a different planning object from a price.

Separate the four risks that get bundled as "repricing", because they have four different owners: a list-price change; an allowance or throttle change; a change to the meter itself, what a credit is worth, which model a request is routed to; and a change in your own volume or model mix. Then instrument five fields from day one, user, tool, task type, credits consumed and the model that served the request, with a cost centre against each, because without task type you cannot kill the workloads that only ever worked while subsidised. And run a tail report before every renewal: a small minority of seats will be a majority of consumption, and knowing that ratio turns a price conversation into a design conversation.

Three items belong in the buyer's model that the market conversation skips: cash, because prepaid credits and committed spend are money out before consumption, and an under-consumed commitment is wasted capacity at best and an impaired prepayment at worst, depending on the contract and the accounting analysis; capitalisation, because most of this is operating expense, but implementation and configuration spend can need a specific analysis under the intangibles and cloud-arrangement guidance, and that needs deciding once; and currency, because a UK company budgeting AI spend is budgeting a dollar cost base, so the translation sensitivity sits next to the repricing one.

Then build for substitution, because that is the mature-phase lever and the one the cloud era eventually forced on every IT budget. Keep the model layer behind an interface you control, so that routing a task to a cheaper model, or to a different provider, is a configuration change rather than a rebuild. The buyer who can switch holds the only leverage left once the product is embedded.

Negotiate the mechanics, not the discount, and start 180 days out: a shadow-billing period before you commit, pooled entitlements, carry-forward of unused credits, tiered rather than punitive overage, protection against changes to the rate card and the credit multipliers, a hard cap as well as an alert, substitution rights, and a reopener if unit prices fall. A carry-forward clause can be worth more over two years than fifteen per cent off list, and none of it exists at the renewal call if it is not in the order form. Then diarise the known dates: Gemini 3.8 Flash doubles on 1 January 2027, GPT-6 Sol arrived on 22 September at $2 and $10, permanent, which retires the GPT-5.6 Sol promotion that ran to 21 November, GitHub's advance seat billing starts with the 1 October cycle, and Microsoft 365 Copilot Business's $18 introductory rate, first year only on an annual commitment, ends on 31 December 2026 against a $21 list.

What to do now, on either side of the invoice

None of this is an argument against adopting AI. The subsidy is real, and for now it buys most buyers capability below what heavy use would cost on a meter; take it, with the repricing on the risk register. Three disciplines.

First, model the allowance and the mix, not a multiple, and note which lever lands on which desk. A price rise lands on the budget and you see it at renewal: procurement owns it. A cut to the quantity inside the plan lands on a delivery date in week eight of the quarter: operations owns it, and it will not appear in the variance you are looking at. In 2026 the second lever moved further than the first. Model both.

Second, instrument the usage now and build for substitution from the start, because when the pricing conversation arrives, and it may arrive as a plans-page update rather than a negotiation, you want to be holding measured value by task rather than anecdote, and an architecture that can route around the price. That is exactly what the O2 machinery did for marketing spend: it turned "we think this is worth it" into "here is the value, by profile". On an interim mandate the first thirty days of this are unglamorous and concrete:

• the five fields switched on, with a cost centre against each;
• a tail report by cost centre, so the renewal is a design conversation;
• a renewal calendar with every promotional expiry on it;
• the contract mechanics listed against each vendor before anyone talks about a discount;
• and a substitution test: which workloads could move to a cheaper model, or another provider, on a configuration change.

If the allowance has already moved on you mid-quarter, the order is different and shorter: triage by task type, throttle the automated cohort first, buy the top-up rather than re-plan the year, and put the renewal date in the diary before anything else.

Third, if your own product bundles AI, do not anchor your customers to a rate you will have to walk back. The operators who trained a generation to expect a free handset spent years unwinding that expectation, at real cost to margin and to trust, with a regulator involved at the end of it; the regulators are already involved at the start of this one. Price for the serving cost you will actually carry, even when a competitor's subsidised rate makes that look expensive today. The first question to put to a board that has quietly built AI into its pricing is this: is the rate one we can sustain, or one we are borrowing against our own future margin?

Adopt now, while someone else pays to build the habit. But adopt with the instrumentation in place and the model built from both sides, so that when the meter appears you meet it on your terms rather than theirs. A subsidy is a bridge. The question is never whether it ends, only when, and what you have built to be standing on the other side.

What to take away

• The quantity inside the seat moved in 2026 more than the rate did, and so did the cash timing; both land on delivery and treasury, not on the price variance.
• Most of AI is early in its growth phase, with the embedded products already behaving like mature ones: the light user, and now the advertiser, is paying for the heavy one, and the netting is breaking as agents replace chat.
• The winners and losers flip when the meter arrives, and the dates are known: GPT-6 Sol at $2 and $10 from 22 September, permanent; GitHub's advance seat billing from 1 October; Copilot Business's $18 to 31 December; Gemini 3.8 Flash doubling on 1 January 2027.
• Model both sides of the invoice, strike acquisition payback on gross profit after serving cost (4.0 months on revenue is 6.6 in the example), instrument the five fields from day one, and build for substitution.
• If your own product bundles AI, price for the serving cost you will carry, not the subsidised one you pay today.

In brief

What actually changed in AI pricing in 2026? The included quantity, not the rate. GitHub cut the credits inside a Copilot seat by more than a third on 1 September 2026 and left the price alone; buyers cut their effective token price by 41 per cent, helped by vendor price cuts and by routing work to cheaper models; and the rate moves are dated in both directions: Gemini 3.8 Flash doubles on 1 January 2027, and on 22 September OpenAI halved its API rates with GPT-6 Sol and Luna while Anthropic released Opus 5.5 at 20 per cent below Opus 5.

Where is AI on the adoption curve, and why does it matter for pricing? Leaving early adoption for growth: light users no longer net off the heavy ones once agents replace chat, so the pricing changes shape, which is what the allowance did in 2026. In maturity the product is embedded as a cost of doing business and the buyer's leverage moves to architecture, whether you can switch what sits behind the interface.

What does "instrument the usage" actually mean? Five fields from day one, user, tool, task type, credits consumed and the model that served the request, each with a cost centre; then a tail report before every renewal, because a small minority of seats will be a majority of consumption.

Why are agents the driver? Because an agent runs a task end to end and bills every step, retry and tool call along the way, many times the cost of a chat exchange. Gartner's August 2026 forecast is a fivefold rise in inference cost per agentic workflow by 2028, and Uber exhausted its full-year 2026 coding-tool budget in four months before capping spend per engineer.

What if my own product includes AI? Price for the serving cost you will genuinely carry, not the subsidised cost you pay today, and design the allowance before the headline. Anchoring customers to an unsustainable flat rate buys adoption you then spend years walking back, which is the trap the mobile subsidy set for the operators who won it.

If AI spend is growing inside your plan and nobody has modelled the day the allowance moves, on either side of the invoice, that is a one-page piece of work worth doing before the vendor does it for you. The model behind the worked example is downloadable here, with the inputs unlocked, and it is a conversation I have directly: get in touch.

Discuss this piece on LinkedIn: linkedin.com/in/jatinderpurewal · More Insights

References

All pricing pages accessed 14 September 2026 and rechecked 18 September 2026; rates change on the vendors' schedules.

• GitHub, "Plans for GitHub Copilot", docs.github.com/en/copilot/get-started/plans - the credit allowances and seat prices as at 14 September 2026 (base credits: Pro $10 with 1,000, Business $19 with 1,900, Enterprise $39 with 3,900; the individual plans add flexible credits on top); GitHub Changelog (1 June 2026), "Updates to GitHub Copilot billing and plans" - the move to AI credits with pay-as-you-go overage; and GitHub Changelog (28 August 2026) - advance payment for Copilot Business and Enterprise seats for card and PayPal customers: new seat assignments before access from September, existing seats from the 1 October 2026 billing cycle, prices unchanged.
• Google (2 September 2026), "Gemini 3.8 Flash" launch post, blog.google - introductory pricing of $0.75 and $3.75 per million tokens through 31 December 2026, then $1.50 and $7.50 from 1 January 2027.
• OpenAI (21 August 2026), "GPT-5.6 Sol" pricing update and the API pricing page, developers.openai.com/api/docs/pricing - the cut to $4 and $20 per million, promotional through at least 21 November 2026, and GPT-6 Astra at $10 and $50 from 3 September 2026.
• Anthropic, pricing documentation, platform.claude.com/docs/en/about-claude/pricing - the cancelled Sonnet 5 increase and the 75 per cent cut to cached-input pricing on Fable 5.1 and Mythos 5.1; Anthropic's developer account announcement of 29 August 2026 (the "17% reduction" wording and the permanent 25 per cent rise) and Anthropic support, "Claude Code May to August 2026 weekly limits promotion" (the 13 May to 13 September window).
• Microsoft, "Microsoft 365 Copilot pricing" - the Copilot Business introductory rate of $18 against $21 list, offer running 1 July to 31 December 2026 on an annual commitment, first year only; and the July 2026 commercial price changes (Office 365 E3 $23 to $26; Microsoft 365 E3 $36 to $39; Microsoft 365 E5 $57 to $60), announced 4 December 2025 and effective at each customer's renewal from 1 July 2026.
• Bain & Company (10 August 2026), Maltiel, E. and Sandberg, J., "AI Pricing: A Reality Check on Effort, Usage, and Outcomes" - four in five vendors introducing AI pricing choosing capacity models; metering split of output about 55 per cent, effort about 35 per cent, outcome about 10 per cent; one in five AI-native companies still mostly per-seat.
• Gartner (17 August 2026), press release, "Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028" - the "inference paradox".
• Ramp Economics Lab (9 September 2026), "AI Index, September 2026" - top one per cent of firms cutting AI spend per employee to $7,205 in August; effective blended price per million tokens down 41 per cent to $0.68 from a March 2026 peak of $1.15; frontier models' share of tokens 45 per cent, down from a 53 per cent peak in August.
• Forbes (17 May 2026), "Uber burns its 2026 AI budget in four months on Claude Code" - the 32 to 84 per cent adoption climb across roughly 5,000 engineers; Fortune (26 May 2026), "Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it's worth it"; TechCrunch (2 June 2026), "Uber caps employee AI spending after blowing through budget in four months" - the $1,500 per employee per month per tool cap.
• Startup Daily (5 August 2026), "Canva cuts revenue forecast by a third as it tackles high AI costs", reporting the Q2 shareholder update first covered by the Australian Financial Review (3 August 2026) - the growth-guidance cut from 30 to 20 per cent and the chief executive's statements on cost to serve; Forbes Australia (19 August 2026), "Canva founder says AI costs down 90% after skyrocketing bills delayed product rollout" - the co-founder's statement on in-house models; Altman, S. (6 January 2025, UTC), post on X: "insane thing: we are currently losing money on openai pro subscriptions! people use it much more than we expected."
• Poyar, K. (13 May 2026), "The State of B2B SaaS and AI Monetization in 2026", Growth Unhinged - hybrid pricing at 37 per cent as primary structure, up from 25 per cent, across 230 B2B software and AI companies; Salesforce Ventures, "AI Pricing: What the Data Says" - more than half of AI sellers on a flat base plus consumption.
• ICONIQ (July 2026), "State of AI: The Builder's Economy" - AI product gross margin of 45 per cent in 2025 and a projected 53 per cent in 2026, from surveys of more than 300 executives; Aleph and Benchmarkit (1 June 2026), "2026 SaaS and AI Performance Benchmarks" - 342 companies, CY2025 actuals: usage-based median net revenue retention 108 per cent against 98 per cent seat-based, median gross retention 84 per cent, median CAC payback 16 months.
• Deloitte (9 September 2026), "Finance Trends 2027: Shaping the next era of stakeholder value" - 1,434 finance leaders in 26 countries: 60 per cent expecting to need more sophisticated AI cost-management practices through 2027; FinOps capabilities at 38 per cent of those preparing for the shift against 25 per cent of those maintaining current practices.
• Competition and Markets Authority (investigation opened 27 July 2026, announced 29 July), "Microsoft: consumer protection enforcement case", gov.uk - the investigation into Copilot's addition to Microsoft 365 consumer plans; Australian Competition and Consumer Commission (27 October 2025), Federal Court proceedings against Microsoft over Microsoft 365 subscription communications.
• Ofcom (22 July 2019), "Helping consumers to get better deals in communications markets: mobile handsets", statement and consultation - the out-of-contract findings (two million bundled customers, around 1.4 million overpaying by just under GBP 11 a month, some GBP 182m a year; more than a quarter better off staying), the voluntary discounts covering nearly 80 per cent of overpaying customers, Three's decision not to apply a discount, the state of split contracts in July 2019, and the fall in bundled contracts from 74 per cent of pay-monthly in 2014 to under half; Oftel (2001), "Effective Competition Review: Mobile" (launched February 2001, published 26 September 2001) - the inclusive-minute structure of UK contract tariffs in 2000-01.
• European Commission (11 May 2016), Case M.7612 Hutchison 3G UK / Telefonica UK, decision, paragraph 792 - O2 Refresh, April 2013, "the first 24 month tariff to decouple the cost of the phone from the cost of mobile services"; OECD (2 July 2013), "Mobile Handset Acquisition Models", OECD Digital Economy Papers No. 224 - the "commonly but inaccurately described as a handset subsidy" framing and "rather a bundled sale"; Bango (19 November 2025), "The rise of the AI subscriber" - almost $66 a month across four tools among 2,000 paying US subscribers.
• OpenAI (February 2026), "Testing ads in ChatGPT", openai.com/index/testing-ads-in-chatgpt, and OpenAI (August 2026), "ChatGPT Ads expands across Europe", openai.com/index/chatgpt-ads-expands-across-europe - ads on the Free and Go plans only, with Plus, Pro, Business and Enterprise ad-free, "advertising doesn't influence the answers ChatGPT generates", and "eight additional markets" between the February US pilot and the European expansion; the 9 February 2026 US start as reported by Ad Age and MediaPost (9-10 February 2026); the UK launch of 6 June 2026, OpenAI's first European market (Canada, Australia and New Zealand had followed the US pilot in March), as reported by Digiday and Mi3; Digiday (31 August 2026), "OpenAI's ChatGPT ads business hits $1 billion run rate as Europe gets self-serve access" - the "$1 billion in ARR in under 200 days" wording from OpenAI's vice-president of global ad solutions.
• Google Ads Help, "About ads and AI Overviews", support.google.com/google-ads/answer/16297775 (rechecked 18 September 2026) - ads inside AI Overviews in twelve countries, in English on mobile and desktop, the UK not listed; ads above or below an Overview in all markets where Overviews run; Google (20 May 2026), "Google Marketing Live 2026", blog.google - the ad formats built with Gemini for AI Mode and Search; Alphabet, first-quarter 2026 earnings call (April 2026) - the chief business officer's remark that a format that works in AI Mode would transfer to the Gemini app.
• OpenAI (22 September 2026), GPT-6 Sol and GPT-6 Luna, API pricing page - $2 and $10, and $0.10 and $0.50 per million tokens, against GPT-5.6 Sol at $4 and $20 and GPT-5.6 Luna at $0.20 and $1.20; VentureBeat (22 September 2026), "OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more" - the spokesperson's confirmation that the rates are permanent; The Next Web (22 September 2026), "OpenAI cuts GPT-6 prices in half with Sol and Luna" - the open-weight competition framing (Alibaba, DeepSeek, Moonshot).
• Anthropic (22 September 2026), "Claude Opus 5.5", anthropic.com/news/claude-opus-5-5, and the pricing page, platform.claude.com/docs/en/about-claude/pricing - $4 and $20 per million tokens against Opus 5's $5 and $25, cache reads $0.20, "costs 40% less to run than Opus 5 on typical workloads".
• Anthropic (4 February 2026), "Claude is a space to think", anthropic.com/news/claude-is-a-space-to-think - "Claude will remain ad-free"; revenue from "enterprise contracts and paid subscriptions".
• Apple Newsroom (9 September 2026), "Apple debuts iPhone 18 Pro and iPhone 18 Pro Max" - the variable-aperture main camera, on sale from 18 September 2026; 9to5Google (21 January 2020), "Samsung Galaxy S20 ditches dual-aperture camera" - the feature carried by the Galaxy S9 and S10 and dropped from the S20.
• Purewal, J. (2026), What AI actually changes in FP&A - the working detail of the AI-in-daily-practice seat this piece speaks from; and The SaaS metric set is the patient funnel wearing different clothes - the subscription-metric mapping the seller-side section leans on; the one-page model: the-meter-under-the-seat-model.xlsx.

The views here are my own. They do not represent the position of any current, former or future employer or client, and nothing here draws on confidential information.