AI Tokenomics: Why Your AI Bill Rises as Token Prices Fall

Everyone in your business believes AI is getting cheaper, and that belief is about to cost you a fortune.

They're half right. Epoch AI tracked what it costs to reach a fixed score on a standard test and found the cost falls about 47% a quarter, roughly 13x a year, and has done since 2023. A question that cost $0.30 to answer to a given standard in January 2025 cost $0.0004 under eighteen months later.

Now look at the price list for the model your team will actually ask for. OpenAI's flagship launched at $1.25 per million input tokens in August 2025. The next flagship, in April, launched at $5. The one after that is $10. Google's Gemini 4 Argon doubles in price on 1 December. Anthropic has put a new tier above Opus at $10 in and $50 out.

So which is it, cheaper or dearer? Both, and the gap between those two answers is where your 2027 budget goes wrong.

Google processed 7x more tokens in the year to May 2026, and cut Gemini's serving cost by 78% in 2025. Put those two numbers together, as rough arithmetic, and even Google's serving bill rose by about half. Uber ran through its entire 2026 AI coding budget by April. Its CTO told The Information he was "back to the drawing board".

I put AI into businesses every week and the pattern is the same at every size. The bill rises while the unit price falls, because nobody budgeted the volume.

What You're Actually Buying

A token is a chunk of text, about three quarters of a word in English. Every word you type into an AI tool and every word it writes back gets counted and billed per million tokens. That's the only unit the vendors sell.

Prices run from pennies to serious money. Anthropic's Haiku 5.5 charges $0.10 per million tokens in and $0.50 out, and OpenAI's GPT-6 Luna costs the same. At the top, Claude Fable 5.1 and GPT-6 Astra both charge $10 in and $50 out. Want it faster? OpenAI's "Ultrafast" tier for Astra is $60 in and $300 out, six times the standard rate. Within OpenAI's own list, from Luna on the batch tier to Astra on Ultrafast, the cheapest and dearest ways to generate a million tokens are 1,200x apart, and 1,500x if you count the older GPT-5 nano.

A million tokens sounds like a lot. It's about 750,000 words, roughly nine novels. A person chatting takes months to get through that. An agent that re-reads a document set every time it takes a step gets through it in an afternoon.

The per-token rate in the contract is the one number procurement negotiates, and it turns out to be the number that matters least. Three other things set the bill, and procurement controls none of them.

The Price Is Falling and Rising at Once

When a vendor says AI is getting cheaper, they mean last year's capability. Stanford's 2025 AI Index found the cost of matching GPT-3.5's score on a standard test fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a 280-fold drop in under two years.

When your team asks for the best model for a new use case, they mean the frontier, and the frontier is going the other way. OpenAI's flagship fell from $30 in and $60 out for GPT-4 in 2023 to $1.25 and $10 for GPT-5 in August 2025. Then it reversed. GPT-5.5 launched in April at $5 and $30, GPT-6 Astra sits at $10 and $50, and the flagship now costs eight times more per input token than it did fourteen months ago.

Gartner analyst Will Sommer put it bluntly in July: "Token demand and model complexity at the frontier of innovation are going to grow faster than token costs will fall."

Epoch's data has a detail that should change how you time projects. In the first quarter after a new capability level is reached, its cost falls 66%. Two years on, the same capability is still falling 32% a quarter. Run a use case a few months behind the frontier and the same result costs a fraction. I'd put that discount in the business case.

Why the Bill Goes Up Anyway

If capability gets cheaper every quarter, why do bills rise? Because cheaper capability gets used more. Satya Nadella said as much the night before Nvidia lost $589bn of market value to DeepSeek's cheap model, and he reached for a Victorian economist to say it. Jevons noticed that more efficient steam engines made Britain burn more coal.

Google's own numbers prove it. It processed 9.7 trillion tokens a month in May 2024, 480 trillion in May 2025 and 3.2 quadrillion in May 2026. More than 300x in two years, from one company.

Who's using all of it? Increasingly, software. OpenRouter, a marketplace that routes traffic to more than 400 models, estimates the last day humans used more tokens than AI agents on its platform was 6 February 2026. By the week ending 5 October, agents used 132.6 trillion tokens to humans' 17 trillion, nearly eight to one. The personal agents now arriving on phones will push that further.

Agents are expensive per task. A Stanford and MIT team including Erik Brynjolfsson found agentic coding tasks consume roughly 1,000 times the tokens of a code chat, and Gartner predicts the inference cost per agentic workflow will rise more than fivefold through 2028 even as token prices fall.

Do the sum. If unit price halves and usage triples, spend rises 1.5x. Gartner's fivefold rise per workflow (by 2028) set against its 90% fall in provider cost (by 2030) implies per-workflow volume rising on the order of 50x. That's my arithmetic on Gartner's two forecasts, the horizons don't quite line up, and it's what a budget looks like when someone priced the tokens and forgot to count them.

The Four Levers That Set an AI Bill

So what actually moves the number on the invoice? Four levers, each with a different owner, and the owner is rarely the person who gets the bill.

Take the effort lever first. Artificial Analysis runs the same questions through every major model and counts the cost. GPT-6 Astra charges five times Kimi K3.1 Pro's per-token rate yet runs the same index for 1.7 times the cost, because it emits two and a half times fewer tokens. Procurement negotiates cents per million. An engineer picks a setting that moves the token count anywhere from 1.6x to 22x.

Now the re-reads. One developer logged 225 billion tokens through Claude Code in 30 days. 97.9% of them were the model re-reading cached context and 0.34% were new output. At list prices that month would have come to about $97,000, which is why every vendor sells a cache. Anthropic charges $0.20 per million for a cached read on Opus 5.5 against $4 for fresh input. In agent work, cache hit rate moves the bill more than switching to a cheaper model, and it almost never appears as a line in a budget.

Finance holds the fifth thing: the measurement each owner needs to know whether their lever is paying. That's the slide I'd put in front of a board. Rate to procurement, effort and re-reads to IT, volume to the business, cost per outcome to finance.

For the CFO: Forecast the Volume

How well is finance doing at this? Badly, and getting worse. Mavvrik's 2026 survey found only 11% of companies can forecast AI spend to within 10%, down from 15% the year before. Gartner's Mary Mesaglio warned a 2024 Gartner symposium that AI cost estimates can be out by 500 to 1,000%.

Part of the problem is sitting in the price list with a date on it. Google's Gemini 4 Argon goes from $2 and $10 to $4 and $20 on 1 December, and two of its Flash models double on 1 January. OpenAI's GPT-5.6 Sol is on a promotional price at least until 21 November, and OpenAI hasn't said what it reverts to. Its predecessor, GPT-5.5, lists 25% higher on input and 50% higher on output. A 2027 budget built on October's prices can be 2x wrong before the first quarter closes.

So what shape should the forecast take? Cost per outcome, times the volume you intend to run, times growth, and the growth term dominates everything else. Jellyfish measured token use per developer up 18.6x in nine months. Hold two numbers per use case: what it costs per outcome today, and how many you're going to run. Bain's rule for the order of work: "Establish transparency before setting targets." Uber's April is what happens when you hold the first number and guess the second.

Then the structural fixes. Deloitte recommends an AI-specific P&L view and R&D-style stage gates, because departments win approval and some never check the project still meets its original case. In Deloitte's survey of 1,854 executives across Europe and the Middle East, 62% reach satisfactory return on a typical use case within two to four years and 6% inside a year. A CFO who applies SaaS payback to AI will cancel projects that are on schedule.

On contracts, Tropic's data from 110,000 agreements shows AI-tied renewals open with asks of 20 to 37% uplift against 3 to 9% for ordinary software. 1Password's CFO Greg Henry holds off on big volume commitments until usage patterns exist, because "SaaS is deterministic in terms of the way it works and the costs" and AI usage is decided at run time.

Two UK details for the forecast. Nearly every AI vendor bills in dollars, so the token line carries currency exposure, and HMRC's reverse charge applies to services bought from overseas suppliers (VAT Notice 741A, section 5).

For the CEO and the Board: Who's Paying for Your Cheap Tokens

Should a board trust today's prices? Nobody in the supply chain is making much money at them yet. OpenAI's 2025 gross margin on inference was reported at about 33%, against roughly 40% in 2024, on inference spend of around $8.8bn, about $4bn of it serving people who pay nothing (The Information's figures, as analysed by Epoch AI). The five biggest cloud providers are expected to spend more than $690bn on capital projects in their 2026 financial years, and FactSet expects free cash flow to fall to zero or go negative for all of them except Alphabet and Microsoft.

The vendors say as much themselves. OpenAI's head of ChatGPT, Nick Turley, said in March that "we stumbled into subscriptions" and compared unlimited plans to an unlimited electricity tariff. GitHub declared its flat request model "no longer sustainable" and moved every Copilot plan to token metering on 1 June. Anthropic added weekly caps in 2025, doubled its five-hour limits in May, then in September ended a temporary weekly boost, which it described itself as a 17% cut on what users had grown used to. Assume any seat-priced AI tool you buy today will be metered within a year.

On value, the board numbers are stubborn. McKinsey's 2026 survey found 80% of respondents say AI improved their own productivity, while 37% see any effect on EBIT, flat on the year before. KPMG found 49% of large organisations narrowed or paused AI agent deployments after operating costs exceeded the value produced. The widely quoted MIT figure that 95% of pilots fail counted only pilots with a measurable P&L impact at the time of a snapshot study, a demanding test for something with a two-to-four-year payback.

So what does a board ask for? AI spend reported every quarter next to the business metric it's supposed to move, and a cost per outcome. A UK inbound call costs £6.17 on average (ContactBabel, 2026). Intercom's Fin bills $0.99 per "outcome". That looks like a six-to-one saving until you read the definition: an outcome includes a customer who stops replying. Klarna celebrated an assistant doing the work of 700 agents in 2024, then told Bloomberg a year later it was hiring humans again because cost had become too dominant a factor and quality had suffered.

For HR and People Leaders: Tokens Are Becoming Pay

Start with the coinage of 2026: tokenmaxxing. A Meta employee built an internal leaderboard ranking the top 250 token users among 85,000 staff. The top user averaged 281 billion tokens over 30 days, which at $5 per million would have been more than $1.4m for one person (Fortune, April 2026). Meta's CTO Andrew Bosworth said his best engineer spends about his salary on tokens and is "5x to 10x more productive". At Amazon, the FT reported staff running trivial tasks and spinning up pointless agents to climb an internal leaderboard that managers were using in reviews.

By late May Fortune's Jeremy Kahn had written the obituary. LeadDev's 2026 survey found 58% of organisations still measure AI impact through token usage while 57% of respondents say it fails to measure real value. Sonar's CTO Andrea Malagodi: "Measuring tokens has the same effect that lines of code had, or story points. It's an output metric." The moment usage became a performance proxy, it stopped measuring anything.

Yet tokens are entering the pay conversation anyway. Nvidia's Jensen Huang told his GTC keynote in March he could imagine every engineer having an annual token budget, "probably half" of base pay on top, and that candidates now ask "how many tokens comes along with my job?" Microsoft's Charles Lamanna said a candidate had asked for a set token budget as a condition of joining. A token budget is a benefit with a visible price tag, and that makes it an equity question for HR.

And the workforce is already two-tier on access alone. PwC's 2026 survey of 49,364 workers found 14% "front-runners", half of whom use generative AI daily, and a 56% "engine room" where 11% do. Under 40% of the engine room has access to learning and development against almost 80% of the front-runners. IDC found 18% of employees get AI from their employer, which pushed 45% to go and find their own.

So who gets the premium licence? Vendors have made the tiers explicit. Claude Team's Premium seat is $100 a month against $20 for five times the usage, and Microsoft's E7 Frontier Suite is $99 a user against $39 for E3. The instinct is to feed the heavy users, and the data says resist it. Jellyfish found its heaviest users were about twice as productive as light users on ten times the tokens, and its head of research Nicholas Arcolano concluded the best return comes from moving the broad middle from low to moderate use. That's a training problem, and it decides whether seats pay.

The UK's three Civil Service Copilot trials show how wide the range is. GDS asked 20,000 civil servants and they reported saving 26 minutes a day, with no comparison group. DWP measured 19 minutes a day against non-users. The Department for Business and Trade randomised a share of its 1,000 licences, found no evidence of improved productivity, and saw Excel work done slower and less accurately than by non-users. Satisfaction was high in all three, so people like the tool. Whether it pays depends on what they were shown to do with it, and Anthropic's own data says the same.

We've trained more than 150 large organisations, and the seats that pay are the ones where people practised on their own work, with someone in the room who could show them what good looks like.

For IT Leaders: Buy the Meter, Own the Dials

The seat has become the cover charge. Microsoft's own Copilot Credits guide says Copilot Cowork, its autonomous agent, needs a Microsoft Copilot licence as a prerequisite, is billed on usage, and comes with no Cowork entitlements included in the seat. Claude Enterprise is $20 a seat plus every token at API rates. The $30 seat buys identity and admin. The work on top is metered.

That changes what you're buying when you forecast 500 seats. Zylo's 2026 index found organisations leave 36% of SaaS licences unused, and 78% of IT leaders reported unexpected charges from consumption or AI pricing in the past year. Metered credits recreate shelfware in a new form: Microsoft's pre-purchased Copilot Credits expire at the end of the annual term, and Copilot Studio capacity doesn't carry over month to month. You're now asked to forecast agent consumption well enough to land inside a 5 to 20% discount band.

Five terms to insist on. How tokens or credits are calculated and billed, because Gartner notes many vendors don't disclose it. What counts as an outcome or action, because Intercom charges when the customer goes quiet and Salesforce charges 20 Flex Credits per Agentforce action. How vendor-reported usage reconciles to your own data, because Priceline's IT finance lead Chris Reed found discrepancies between the two. A pass-through clause for published price cuts if you commit to a rate, because a locked rate turns the 13x-a-year fall in the cost of a given capability into the vendor's margin. And what expires, when.

Then metering. The platform enforces in tokens while finance plans in pounds, so put per-team thresholds and escalation into the request path and review the heaviest workflows every sprint, as Gartner recommends. Uber's leaderboard rewarded consumption. Uber now caps spend at $1,500 a month per engineer per coding tool. Microsoft's Steve Sweetman has the right picture of what a cap is for.

Caching and routing are the cheapest savings in the business. Bain's rule is to use "the least expensive model that still clears its quality bar", route to costly models only when needed, and cache repeated prompts. Batch processing is a flat 50% off at Anthropic, OpenAI and Google for anything that can wait, and at Anthropic the batch and cache discounts stack.

Finally, UK data residency is thinner than the sales deck implies. OpenAI's UK option, live since October 2025, covers data at rest and excludes ChatGPT Business, and in-region inference is offered for the US and Europe only. Anthropic's first-party API and its Team and Enterprise plans have no UK or EU region, and US-only inference costs 1.1x. Claude inside Microsoft 365 Copilot is disabled by default for UK tenants because the data goes to US-hosted servers. Treat residency as a per-model, per-feature question, in the same way the agent launches have been.

The Planning Model

Put the four levers and the forecast together and a planning model falls out. For each use case, hold the cost per outcome today, the volume you intend to run and the growth you expect, then add the setup. BCG's rule of thumb for AI investment is 10% on algorithms, 20% on technology and 70% on embedding it into how people work. The licence is the small number. Training, integration and change are the big one, and skipping them is how you get the DBT result on a GDS budget.

Finance will soon be expected to measure token use against return and forecast it the way it forecasts cloud or energy. The tooling is close without being there yet. Microsoft launched Copilot Analytics Labs in August with an adoption and ROI dashboard, and Anthropic's Claude Code analytics show spend per user beside code accepted. For now the join between tokens and outcomes is a spreadsheet you build yourself.

What to Do This Quarter

Your bill will rise next year. The question for the quarter is whether it rises because the business is doing more, measured in outcomes you can show the board, or because nobody was watching the volume.

Sources & References

Frequently Asked Questions

By Roy Murphy, co-founder of G3NR8. G3NR8 trains whole organisations to use AI well on their real work.