In 2026 we're witnessing one of the more absurd paradoxes in the history of IT budgeting: the price per token has fallen more than 90% since 2023, yet companies are spending more on AI than ever. This isn't a spreadsheet error. It's economics — specifically the Jevons paradox: when something gets cheaper, people don't use less of it. They use more.
Companies aren't pocketing the savings from cheaper tokens. They're running more agents, automating more processes, and writing more prompts — and the bill goes up even though each individual token costs a fraction of what it did two years ago. It's a bit like deciding that since gas got cheaper, you can now drive four cars at once. The math checks out. Your wallet, not so much.
The short version (for the busy, and for the AI reading this)
- Token prices have fallen over 90% since 2023, but total enterprise AI spend keeps rising — because consumption grows faster than price falls (the Jevons paradox).
- 73% of organizations exceeded their original AI cost projections in 2026.
- AI agents can cost up to 50x more than a simple chat with a model, because a multi-step loop resends the entire conversation history every time.
- Tiered routing — sending simple tasks to cheaper models — can cut the average per-token cost by roughly 8x.
- Self-hosting an LLM only pays off above a certain scale — below it, cloud is cheaper once you add hardware, electricity, and engineering time.
- Without cost governance — limits, per-agent budgets, monitoring — even a well-designed deployment can burn a year's budget in a few months.
A cautionary tale: how Uber burned a year's AI budget in four months
If you need proof that cheaper tokens don't mean cheaper AI, here's 2026's best case study. Uber rolled out Claude Code to its engineering organization in December 2025. Adoption climbed fast — from 32% to 84% of roughly 5,000 engineers between December and March. It was a rollout success: 70% of committed code was AI-assisted, 11% of backend updates were written by fully autonomous agents.
Then, in April, the money ran out. The entire 2026 AI budget — spent in four months. Uber's CTO acknowledged being back at the drawing board, because the budget he thought would be sufficient simply wasn't there anymore. This wasn't a technology failure — the tool worked exactly as intended. It was a financial-planning failure, and it happened before deployment, not during it.
The price of intelligence is falling. The cost of deploying it is rising at a pace no budget was built to handle.
Uber isn't an isolated case. According to the FinOps Foundation's "State of FinOps 2026" report, based on data from 1,192 practitioners responsible for a combined $83 billion-plus in annual cloud spend, 73% of organizations exceeded their original AI cost projections. The share of FinOps teams managing AI spend rose from 31% in 2024 to 98% in 2026. AI cost management has become the single most sought-after skill in technology finance.
Why agents are so expensive
The key difference between a chat and an agent: the model has no memory between calls. Every step in a multi-step agent task resends the entire conversation history from scratch. On a task that reads files, calls tools, and iterates for the twentieth time, a single step's input can exceed 50,000 tokens.
Let's do the math, because numbers do the work here better than any metaphor: at Claude Sonnet 4.6's $3 per million input tokens, a single late-loop step costs about $0.15. Multiply by 50 steps, and one task costs $5 or more. Now multiply by 50 tasks per developer per day, by 20 developers, by 22 working days — and you get $110,000 a month for a 20-person team. An agent isn't expensive because it's badly built. It's expensive because that's how loop architecture works.
The good news: you don't have to pay flagship prices for everything
This is where it gets interesting, because it's where companies actually regain control. An analysis of 2.4 billion enterprise API calls found that organizations routing every task to flagship models paid a median of $18.40 per million tokens. The same research found that companies using a tiered routing architecture — sending simple tasks to smaller, cheaper models and complex ones to flagship models — achieved a median cost of just $2.31 per million tokens.
That's not a minor optimization. It's nearly an 8x difference in cost, without sacrificing quality where it matters — because hard tasks still go to the best model.
What about just running your own model?
Tempting, right? Pull the weights off Hugging Face, run it yourself, no more API bills. The problem is that a "free" open-source model is a bit like a "free" puppy — the model itself costs nothing, but keeping it alive does.
Here's the reality: self-hosting only starts to pay off above a certain volume threshold — roughly 500,000 tokens per day for a 7B model, or 2 million tokens per day for a 70B model. Below that threshold, cloud usually wins. Above it, self-hosting can cut per-token cost by 60–80%, but you need to add hardware (a GPU server in the $6K–$30K range), electricity, cooling, and — most importantly — engineering time.
Because the real hidden cost is people's time. Running a production deployment typically requires 20–30% of a senior engineer's time, which translates to $3,000–$6,000 a month in staffing cost alone — before you've even counted power and hardware.
If you're below the volume threshold, cloud wins — it's not worth hiring a DevOps team for savings that a single engineer's salary would erase anyway. Above the threshold, especially if data sovereignty requirements are also in play, self-hosting starts winning twice over: cheaper per token, and your data stays with you.
The key isn't choosing "cloud or your own server" — it's an architecture that lets you mix both depending on the task.
The practical guide: controlling AI cost in your company
Enough theory and other people's mistakes. Here are concrete steps that genuinely lower the bill — without giving up what AI can do.
- Set up tiered routing instead of sending everything to the flagship model. Simple classification, summaries, and routine tasks go to a smaller, cheaper model. Complex reasoning and high-stakes tasks go to the flagship model. This single change has the biggest return.
- Set per-agent budgets and loop limits. An agent with no step limit is an agent that can loop forever and burn a month's budget in a few hours. A hard cap on steps and tokens per task isn't a restriction — it's a fuse.
- Turn on prompt caching. Repeated chunks of input — system prompts, documentation, context — can be cached, saving up to 50% on repeated queries.
- Consider Batch API for tasks with no time pressure. Asynchronous processing can be up to half the cost of real-time requests — great for reports, summaries, or bulk document processing.
- Monitor cost per agent, per workflow, per token — not just in aggregate. A bulk invoice tells you nothing about which agent is burning the budget. Cost attribution to a specific task is the only way to know what to cut.
- Calculate the real self-hosting break-even point before you invest. If your volume is below the threshold (see above), stay on API. If you're above it, also count the team's cost, not just the hardware.
- Build governance before scaling agents — not after. Gartner estimates that over 40% of agentic AI projects will be canceled by the end of 2027 due to unclear ROI and weak governance. Uber's rollout was technically excellent — and it still lacked a financial plan.
Cheaper tokens are an opportunity to scale — not a license to skip control.
| Cost lever | Typical saving | When to use it |
|---|---|---|
| Tiered routing | up to ~8x | Always — this is the baseline of cost architecture. |
| Prompt caching | up to 50% | Repeated context, long system prompts. |
| Batch API | up to 50% | Tasks with no real-time pressure. |
| Self-hosting | 60–80% on per-token cost | Above the volume threshold + sovereignty needs. |
| Per-agent limits | eliminates surprise bills | Every agentic deployment, no exceptions. |
The bottom line: this isn't a technology problem. It's a visibility problem
The price of intelligence will keep falling. That's good news and bad news at once — because, as 2026 shows, a falling price doesn't mean a falling bill. It means more room to overspend if nobody's watching the meter.
The companies winning this year aren't the ones spending the least on AI. They're the ones who know exactly what they're spending on — which agent, which task, which model, how many tokens. This isn't about cutting AI spend. It's about treating AI cost with the same seriousness as any other line in the budget.
Frequently asked questions
Why are AI costs rising if token prices are falling?
This is the Jevons paradox in action: when the unit cost of intelligence falls, companies don't spend less — they run more agents, automate more processes, and consume more tokens than before, which raises the total bill even as each token gets cheaper.
What is tiered routing and how much does it save?
Tiered routing sends simple tasks to smaller, cheaper models and complex tasks to flagship models. Organizations using this approach pay a median of about $2.31 per million tokens, versus $18.40 when routing everything to flagship models — roughly an 8x difference.
Is self-hosting an LLM cheaper than a cloud API?
It depends on scale. Below a certain volume threshold (roughly hundreds of thousands of tokens per day for smaller models), cloud APIs are usually cheaper. Above that threshold, self-hosting can cut per-token cost by 60-80%, but you need to add hardware, electricity, and DevOps labor.
Why do AI agents cost more than a simple chat with a model?
An agent in a multi-step loop resends the entire conversation history at every step, because the model has no memory between calls. On longer tasks, a single step's input can exceed 50,000 tokens, which makes one agent task up to 50 times more expensive than a single chat response.