The price of an AI token has fallen something like 80-90% in two years. Your company’s AI bill has almost certainly gone up anyway. That is not a billing error, it is the central paradox of enterprise AI in 2026: the cost of intelligence is dropping while the cost of using it climbs. The reason is simple. You are using far more tokens per task than you used to, especially with agents and bloated context. The good news is the fix is mostly architecture, not magic, and well-run teams are cutting bills 60-80% without losing quality. Here is why it happens and exactly what to do.
Here is a number that should make any operator do a double take. Token prices have fallen by roughly 80-90% over the past two years, depending on the model. By that logic, your AI spend should be shrinking. For almost everyone, it is doing the opposite.
This came up on a recent Bricks & Bytes panel, where Owen Drury pointed out something he keeps hearing across the industry: despite cheaper tokens, actual enterprise AI bills keep going up. People are simply using AI more, and in construction specifically, often using it badly. Upload a 50-page PDF to answer one small question and the model has to read all 50 pages to give you the answer. That is a lot of tokens for a one-line response.
The pattern is now well-documented beyond the podcast. Gartner forecasts that by 2030, running a query on a top-tier model will cost AI firms nearly 90% less than it did in 2025, yet warns that cheaper tokens will not mean cheaper enterprise AI, because modern AI burns far more tokens per task. The unit got cheaper. The bill got bigger. Both are true at once, and the gap between them is the whole story.
Falling token prices are real but irrelevant if your usage grows faster. Agentic workflows and bloated context can multiply token consumption per task by orders of magnitude. The companies winning here are not the ones with the cheapest model, they are the ones with the tightest token discipline.
Why the bill keeps climbing
The unit got cheap, the usage exploded
The core mechanism is volume. When AI meant one prompt and one response, costs were predictable. Today a moderately complex agentic workflow, say researching a topic, drafting an output, checking it against a policy, then revising it, can consume anywhere from 50,000 to 500,000 tokens before it hands you a single answer. The per-token price fell, but the number of tokens per task multiplied far faster. Net result, the bill rises.
And the people approving the spend often do not see it coming, because the number they watch, price per million tokens, keeps falling. Adoption scales faster than governance. Usage grows first, visibility shows up later, if at all. By the time finance asks what happened, the agents have been quietly chewing through context for months.
This is not a small-company problem dressed up as a big one. Uber’s CTO told The Information that the company burned through its entire 2026 AI coding tools budget in just four months, after actively pushing adoption with internal leaderboards. Goldman Sachs projects agentic AI could drive a 24-fold increase in token consumption by 2030. The direction of travel is clear, and it points up.
The cost of intelligence is falling. The cost of deploying intelligence is skyrocketing. Understanding why is now the most important AI financial discipline there is.Industry analysis on inference economics, 2026
The construction-specific trap
Context bloat is quietly the biggest waste in AEC
Owen’s 50-page-PDF example is not a throwaway line, it is the single most common way construction firms waste money on AI. The industry runs on enormous documents: specs, drawings, contracts, RFIs, submittals. The instinct is to dump the whole thing into the model and ask a question. But every page you send is input tokens you pay for, on every single query, whether the answer needed page 2 or page 47.
Do that across a team, all day, and the bill balloons for output you could have gotten from a fraction of the context. Worse, stuffing huge documents into the prompt does not reliably improve the answer. Past a point it can muddy it. You are paying more to potentially get a vaguer response. That is the opposite of what anyone wants.
The fix is not to use AI less. It is to feed it less, more precisely. Retrieve the relevant section instead of the whole document. Route the simple question to a cheap model and save the expensive one for genuine reasoning. None of this requires a data science team, just a bit of discipline about what goes into the model in the first place.
- The whole-document dump. Uploading a 50-page spec to answer a one-line question means paying to process 50 pages every time.
- One model for everything. Sending a simple data-extraction task to a premium reasoning model can cost many times more than a budget model that would do it just as well.
- No reuse. Re-sending the same standard context on every query instead of caching it, when caching can cost a fraction of the normal input rate.
Spend smarter on construction tech
We break down the tools, costs, and traps that actually matter for AEC, once a week, no fluff.
Join 2500+ ReadersHow to actually fix it
Four levers, in rough order of impact
The encouraging part is that this is a solved problem if you put a little effort in. Teams that apply the right combination of techniques routinely report cutting AI costs by 60-80% without degrading output quality. The biggest single lever is usually model routing: instead of sending every query to your most expensive model, a routing layer picks the cheapest model that can actually handle the task. One worked example sent 70% of queries to a budget model, 20% to a mid-tier model, and 10% to a premium model, and cut the bill by about 75%.
The second lever is context management, exactly the cure for Owen’s PDF problem. Retrieve only the relevant chunk rather than dumping the whole file, and you can cut token usage on those tasks by 20-40% straight off. The third is caching, where frequently reused context, like a standard system prompt or a company policy, is stored and billed at a small fraction of the normal rate, with some providers charging as little as 10% of base input price for cached content.
The fourth lever is the one nobody likes but everybody needs: visibility and governance. You cannot manage what you cannot see. Most runaway bills happen because no one was watching consumption per team or per task until the invoice landed. One caution though. Do not compress or trim blindly. Aggressive optimization without checking output quality can quietly erode the answers your people rely on, so measure quality before and after every change.
| Lever | What it does | Typical saving | Effort | Best first move | Watch out for |
|---|---|---|---|---|---|
| Model routing | Sends each task to the cheapest capable model | 20-60%+ | Medium | Send simple extraction to a budget model | Routing complex tasks to a weak model |
| Context management | Feeds only relevant text, not whole documents | 20-40% | Low to medium | Retrieve the right section, not the full PDF | Trimming so hard you lose needed detail |
| Caching | Reuses stored context at a fraction of the rate | Up to ~90% on cached input | Low | Cache standard system prompts and policies | Caching content that changes often |
| Visibility and governance | Tracks spend per team, task, and project | Indirect but large | Low | Set per-project budgets and dashboards | Setting up controls after the bill, not before |
Because the price per token fell, but the number of tokens you use per task rose much faster. Token prices have dropped roughly 80-90% in two years, yet most organizations now spend more on AI than ever, driven by heavier usage, agentic workflows, and bloated context. The unit cost and the total bill are moving in opposite directions. (Source)
Not on their own, according to Gartner. It forecasts inference on a top-tier model will cost AI firms nearly 90% less by 2030 than in 2025, but warns this will not translate to cheaper enterprise AI, because agentic models need far more tokens per task, rising consumption can outpace falling unit prices, and providers will not fully pass on the savings. (Source)
A lot more than people expect. A moderately complex agentic workflow, researching, drafting, validating against a policy, then revising, can consume anywhere from 50,000 to 500,000 tokens before returning a single result. That is why per-task cost, not per-token price, is the number that matters for budgeting. (Source)
Model routing is a layer that sends each request to the cheapest model capable of handling it, rather than defaulting every query to your most expensive model. In production it typically cuts token costs by 20-60%. One example routing most queries to a budget model and reserving a premium model for the hardest 10% reduced spend by about 75%. (Source)
Every page you send is input tokens you pay for, on every query. Dumping a 50-page document to answer one question means the model processes all 50 pages each time. Context management, retrieving only the relevant section, can cut token use on those tasks by 20-40%, and it often improves answer quality rather than hurting it. (Source)
Start with the two cheapest, highest-impact moves: stop dumping whole documents into prompts, and route simple tasks to cheaper models. Add caching for standard reused context, then put basic visibility in place so you can see spend per team or project. Measure output quality before and after each change so you do not save money by quietly degrading answers. (Source)
Indexnine: The token cost illusion
Fortune: AI’s real cost problem and the Gartner forecast
Iternal: LLM pricing and model routing example
Redis: LLM token optimization techniques
Obvious Works: Saving up to 80% on LLM costs
Related Articles
Nemetschek Acquires HCSS: Why the Infrastructure Bet Is the Biggest Signal in ConTech Right Now
The Bricks & Bytes Weekly Newsletter
