finance
Your AI Bill Is Lying to You
Token Economics Series: Part 1: Your AI Bill Is Lying to You | Part 2: The Practitioner’s Toolkit | Part 3: Scaling the Discipline | Cheat-Sheet
New AI models are arriving that are dramatically more powerful — and dramatically more expensive. Most people are already burning tokens inefficiently and don’t know it. Token management isn’t an afterthought or an admin task — it’s a core professional skill, as fundamental as knowing your tools. This series lays out the discipline: why it matters, how to build it, and how to scale it across a team.
The Billion-Token Brag
I keep seeing it. On LinkedIn, on X, in Slack channels, in conference talks. People proudly sharing their token consumption like it’s a high score.
“Burned through 800 million tokens last month.” “Our team hit a billion tokens in Q1.” “I’m spending $3,000 a month on API calls — this is the future.”
And everyone nods along. Impressed. Envious, even. Like the number itself proves something.
It doesn’t.
Here’s a parallel that software people should find uncomfortable. Imagine someone at a conference saying: “I wrote 37,000 lines of code last sprint.” You wouldn’t be impressed. You’d be concerned. 37,000 lines isn’t a badge of honour. It’s a code smell. It tells you nothing about what was achieved and quite a lot about how it was achieved. The interesting question isn’t how many lines — it’s what those lines do, and whether you could have done the same thing in 4,000.
A billion tokens is the same kind of empty metric. It tells you someone used a lot of AI. It tells you nothing about whether they used it well.
I’ve watched people pump entire codebases into context windows for a simple question. I’ve watched teams run the most expensive model available for tasks that a model ten times cheaper would handle identically. I’ve watched developers paste the same 50,000-token document into every prompt in a session because they couldn’t be bothered to figure out which parts were actually relevant. I’ve done all of these things myself.
The distinction matters: were those clever tokens or inefficient slop?
Clever tokens are tokens that did work. They carried necessary context. They powered a model that needed that capability for that specific task. They contributed to output you actually used. Every token earned its place.
Inefficient slop is everything else. The context you stuffed in because “more is better.” The premium model you used for a task that needed a classifier. The massive prompt you ran seventeen times because you didn’t structure it to get the answer in three. The conversation that ballooned to 200,000 tokens because you never started a fresh context window.
Nobody tracks this. Nobody even thinks about it. The culture right now is: tokens are cheap, use more, move fast. And the people spending the most are the ones getting the most respect.
That’s about to become a very expensive mistake.
The Storm Coming
Let’s talk about what’s happening in the next twelve to eighteen months, because the current “tokens are cheap, who cares” attitude is about to collide with reality.
The models arriving in 2026 and 2027 are significantly more powerful than what we have now. Longer context windows. Better reasoning. More reliable tool use. Multimodal capabilities that actually work. The jump from where we are to where we’re going is substantial — and I say that as someone who’s been building with these tools daily for over a year.
They’re also significantly more expensive. The frontier models — the ones that will unlock the use cases everyone’s excited about — are pricing at multiples of what the current generation costs. Not small multiples. Reasoning models that “think” before responding are already burning through tokens at rates that make standard models look like rounding errors. Extended thinking, chain-of-thought, multi-step verification — all of these consume tokens that you pay for but never see in the output.
And it’s not just model cost. It’s model choice. A year ago, most teams used one model for everything. Today, you could credibly use five different models for five different tasks — and you probably should, because the cost and capability spread across available models is enormous. We found cost differences of up to 95% for the same classification task across different models. Ninety-five percent. That’s not optimisation — that’s the difference between a viable product and one that bleeds money.
More models. More techniques. More data flowing through inference. More capability driving more usage driving more cost. This is a compounding problem — and it compounds in both directions simultaneously. The models get more capable, so you want to use them for more things. You use them for more things, so your bill grows. Your bill grows, so you need to be smarter about how you use them. But being smarter takes effort, and the new capabilities are so shiny that nobody wants to slow down long enough to think about efficiency.
Here’s the part that really gets me: reasoning tokens. The latest generation of models don’t just produce output — they think first. Chain-of-thought, self-verification, multi-step planning. All of that thinking burns tokens. Tokens you pay for. Tokens you never see in the final output. Your visible response might be 500 tokens, but the model burned 8,000 tokens of internal reasoning to produce it. If you don’t understand this mechanic, your cost estimates will be wrong by an order of magnitude. And this is the direction of travel — future models will do more thinking, not less.
Jevons Paradox applies here too. When you make something more efficient, people use more of it — not less. As AI gets more capable, teams will use it for more tasks, with more data, in more workflows. The total spend goes up even if the per-unit cost goes down. And for many of the most powerful new capabilities, the per-unit cost is going up.
The industry sees this coming. Every major provider is investing heavily in inference infrastructure — not because demand is slowing, but because it’s about to explode. The economics of AI inference are becoming a first-order business concern, not an IT line item you glance at quarterly.
If you don’t have a discipline for managing token consumption now — when models are relatively cheap and use cases are relatively simple — you are going to be in serious trouble when the next generation lands and your team’s appetite for AI has tripled.
FinOps Had This Problem Already
This isn’t the first time the tech industry has gone through this cycle. Cloud computing followed the exact same arc.
Phase one: it’s easy to provision. Spin up a server in minutes instead of weeks. Everyone’s excited. Usage explodes.
Phase two: it’s hard to manage. Servers running that nobody remembers starting. Development environments that cost as much as production. Storage buckets growing without limit because nobody set a lifecycle policy. The bill arrives and everyone’s surprised.
Phase three: it’s expensive to ignore. A startup discovers they’re spending $400,000 a year on AWS and can’t explain where half of it goes. An enterprise finds that 40% of their cloud resources are idle. The CFO starts asking questions that engineering can’t answer.
Phase four: a discipline emerges. FinOps — the practice of bringing financial accountability to cloud spending. Not restriction. Not austerity. Intentional spending. Knowing what you’re spending, why you’re spending it, and whether you’re getting value for it.
The FinOps Foundation built a maturity model that maps this journey: from “Crawl” (basic visibility into what you’re spending) through “Walk” (active management and optimisation) to “Run” (automated governance and real-time decision-making). Most organisations that went through this journey report that the savings weren’t the biggest win — it was the clarity. Knowing what things cost changed how people made decisions. Teams that understood their cloud bill made better architectural choices. Not cheaper choices. Better ones.
AI inference is at exactly the same inflection point. But it’s moving faster.
Cloud took a decade to go from “spin up whatever you need” to “we need FinOps.” AI inference is going to compress that into two or three years, because the spend is scaling faster, the options are more complex, and the technology is changing underneath you while you’re trying to manage it.
We’re somewhere between phase one and phase two right now. Most teams are in the “it’s easy to provision” stage — sign up for an API key, start making calls, worry about the bill later. Some are hitting phase two — the bill is getting uncomfortable, but nobody has a systematic approach to managing it. Almost nobody is at phase three or four.
The key insight from FinOps — the one that matters most for what we’re building here — is this: it’s not about spending less. It’s about spending deliberately. The teams that got FinOps right didn’t cut their cloud bill to zero. They often didn’t even cut it dramatically. They made it legible. They knew what every dollar bought them. And that knowledge — that visibility and intentionality — changed everything downstream.
That’s what token discipline is. Not austerity. Intentionality.
What Token Discipline Actually Looks Like
Let me be specific about what I mean, because “token discipline” can sound like “use AI less,” and that’s not what I’m saying at all. I’m saying: match the tool to the task. Every time.
The first lever — and the one with the biggest immediate impact — is model selection. Most people use the same model for everything. Their coding assistant, their analysis tasks, their classification jobs, their creative work — all running through whatever model they defaulted to on day one. This is like using a sledgehammer for every task in a workshop. Sometimes you need a sledgehammer. Usually you need a screwdriver.
We wrote about how we’re choosing inference models for different tasks within a single product. The short version: our classification tasks can run on models that cost a fraction of what our creative tasks require. Using the frontier model for classification is like hiring a surgeon to put on a plaster. The surgeon can do it. But you’re paying surgeon rates for a plaster-level task, and the surgeon has better things to do.
Model selection alone — just picking the right model for each task — can reduce token costs by 50-90% without any loss in output quality. That’s not a typo. The spread between models is that wide.
The second lever is context management — and this is the silent cost driver that almost nobody pays attention to.
Every token in your context window costs money. Input tokens aren’t free. And most people are stuffing their context windows with material they don’t need. Entire files when they need three functions. Complete conversation histories when the last two exchanges are the only ones that matter. System prompts the size of novellas when a focused paragraph would produce the same result.
Context bloat is invisible because you don’t see it. You see the output. You see the response time. You don’t see the 80,000 tokens of context you’re paying for on every single request, of which 60,000 are contributing nothing.
The third lever is prompt engineering — and I don’t mean the “write better prompts” advice that fills every AI newsletter. I mean structural efficiency. How many round trips does it take to get what you need? Are you asking for the right thing the first time, or are you refining through five iterations because your initial prompt was vague? Are you using the model’s capabilities — structured output, specific instructions, examples — or are you hoping it reads your mind?
A well-structured prompt that gets the answer in one shot costs a fraction of a vague prompt that takes five attempts. And the five-attempt version isn’t just more expensive — it’s slower, more frustrating, and produces worse results because each iteration introduces drift.
These three levers — model selection, context management, prompt engineering — form the foundation. But they’re individual skills. To make this work at team scale, you need a framework.
Here’s the skill stack we use, and it’s what this series unpacks:
Visibility. You can’t manage what you can’t see. Know what you’re spending, per task, per model, per team member. This is the “Crawl” phase — just getting the data.
Control. Once you can see it, start making intentional choices. Model routing. Context budgets. Prompt templates. This is where individual discipline starts producing measurable results.
Architecture. Build the infrastructure that makes good choices the default. Model routers. Cost tracking. Caching layers. Fallback chains. This is where the discipline becomes systematic rather than heroic.
Culture. Make token awareness part of how your team thinks about AI work. Not as a constraint — as a quality signal. The team that thinks about token efficiency is the team that thinks about what they’re actually doing with AI, and that team builds better things.
Visibility, control, architecture, culture. That’s the progression. You can’t skip steps — culture without visibility is just wishful thinking, and architecture without control is premature optimisation.
The Cost of Not Caring
“Fine,” you might say. “But tokens are still relatively cheap. Why should I care about this now?”
Three reasons, and the first one is the one that will hit you this week, not this quarter.
Rate limits. Everyone who works with AI has hit rate limits. The request that fails because you’ve exceeded your quota. The coding session that grinds to a halt because you’ve burned through your allocation. The demo that dies in front of a client because you didn’t realise how close you were to the ceiling.
Most rate limit problems are self-inflicted. Not because you’re doing too much work — because you’re doing the same work too expensively. Stuffing 100,000 tokens of context into requests that need 10,000. Using a model that generates 4,000 tokens of output when you need 400. Running the same expensive prompt repeatedly because you didn’t cache the result or structure the workflow to avoid redundancy.
Cut your per-request token consumption by 60% — which is entirely achievable with basic discipline — and your rate limit problems disappear. Not because you got a higher limit. Because you stopped wasting the one you had.
Financial waste. This one scales. An individual wasting $50 a month in unnecessary tokens is annoying but survivable. A team of twenty, each wasting the same proportion, is $12,000 a year. An organisation with hundreds of AI users, no consumption discipline, and models getting more expensive? That’s a budget line item that’s going to trigger an uncomfortable conversation with finance. And when that conversation happens, the response is usually a blunt cost cut — reduced access, cheaper models, usage caps — rather than the targeted optimisation that would actually solve the problem without hurting productivity.
Slower iteration. This is the one people underestimate. Bloated context doesn’t just cost more money — it makes everything slower. Larger contexts mean longer response times. Longer response times mean fewer iterations per hour. Fewer iterations mean less experimentation, less refinement, less learning. The developer who runs lean, focused prompts gets ten iterations in the time it takes the context-stuffer to get four. Over a day, that’s a significant productivity gap. Over a month, it’s a competitive advantage.
And then there’s the compound cost — the one that connects back to everything we’ve written about at Viewyonder. Entropy at velocity described how AI-assisted development generates code faster than humans can track, leading to codebase drift and rework. Every wasted token in that cycle makes it worse. Bloated contexts produce less focused output. Less focused output requires more correction. More correction burns more tokens. The rework cycle is a token furnace, and most people are stoking it without realising.
The competitive angle is real too. Teams that develop token discipline now — while it’s still a differentiator rather than table stakes — will build better products, iterate faster, and scale their AI usage more effectively than teams that treat it as an afterthought. When the more powerful (and more expensive) models arrive, disciplined teams will adopt them strategically while undisciplined teams will either overspend or under-adopt. Neither is a good outcome.
Think about what happens when two teams are building competing products and both have access to the same models. Same APIs. Same capabilities. Same pricing. The team with token discipline gets more iterations per dollar, hits rate limits less often, ships faster, and has the headroom to use premium models where they actually matter. The team without discipline burns budget on waste, gets throttled at critical moments, and eventually gets told by finance to cut back. The disciplined team scales up. The undisciplined team scales down. Same tools. Different outcomes. The difference is the skill, not the spend.
Where We Go From Here
This article is the “why.” The provocation. The argument that token management deserves to be taken seriously as a professional discipline, not dismissed as penny-pinching or premature optimisation.
In Part 2: The Practitioner’s Toolkit, we get practical. Model selection frameworks. Context management techniques. Prompt efficiency patterns. The specific tools and habits that make a measurable difference, starting with what you can do tomorrow.
In Part 3: Scaling the Discipline, we zoom out. How do you build token awareness into a team? Into an organisation? What does the infrastructure look like? What does the culture look like? How do you move from individual discipline to systematic practice?
And the Cheat-Sheet distils the whole series into a single reference you can pin to your wall — or, more realistically, your Notion sidebar.
The question isn’t whether you can afford to use these models. It’s whether you can afford to use them badly.
Steve Chambers is the founder of Viewyonder, a GenAI consultancy pioneering people with AI at work. He spends his days helping teams build AI-powered products — and watching them burn through tokens like there’s no tomorrow. If your team’s AI bill is climbing and you’re not sure what you’re getting for it, get in touch.