Here is the answer most vendors will not lead with: the model call is the cheapest part of an AI project, and it is getting cheaper every year. If you are pricing your first serious AI effort by staring at a per-token rate, you are watching the one line on the invoice that barely matters. The budget lives somewhere less glamorous: the integration, the data work, the evaluation that tells you the thing actually works, and the upkeep that runs forever. Price the project on that, and you will plan for the right number. Price it on the token meter, and you will be surprised twice, first by how little the calls cost, then by how much everything around them does.
The token bill is real, small, and shrinking
Start with the part everyone fixates on, so we can set it aside honestly. As of June 2026, a capable workhorse model runs about a dollar per million input tokens and five dollars per million output tokens (Anthropic’s Claude Haiku 4.5 is priced at exactly that on the company’s own pricing page, read June 27, 2026; these are perishable numbers and they move). A smaller model is cheaper still: OpenAI’s GPT-5.4-nano is listed at twenty cents per million input tokens and $1.25 per million output, on OpenAI’s own page the same day. A million tokens is roughly 750,000 words, so for the price of a sandwich you can process the equivalent of several novels. Both of those are provider-published prices, which means they are also marketing, so verify them live before you model anything.
And the direction is unusually clear. Stanford’s AI Index found the cost of querying a model at GPT-3.5’s level fell from $20.00 per million tokens in late 2022 to $0.07 by October 2024, a roughly 280-fold drop in about eighteen months. Two independent instruments point the same way: Epoch AI, regressing prices across six benchmarks, puts the median decline near 50x per year (faster, around 200x, after January 2024); and a16z’s widely-cited “LLMflation” analysis pegs it near 10x per year for equivalent performance, with a GPT-3-quality model falling about a thousandfold in three years. Read those together and hedge the exact multiple, because they disagree by task and method. The robust claim is the one they share: for a fixed level of capability, inference gets dramatically cheaper, fast. That is a tailwind. It is also the trap, because it makes the cheap thing feel like the whole thing.
The costs that actually move the budget
What you are really buying is not access to a model. It is a working system wrapped around one, and the wrapping is where the money goes. Four lines dominate, and none of them is the token bill.
Integration. The model has to reach your data and your tools, return results where work happens, and handle the cases that break. Connecting it to your CRM, your ticketing system, or your document store, and making that connection reliable, is ordinary software engineering, billed at ordinary software-engineering rates.
Data work. Most AI value depends on the model seeing your information, which means collecting, cleaning, and structuring it first. This is the quiet majority of the effort and the most underestimated. By Anaconda’s survey, data scientists spend roughly 45 percent of their time just preparing data (the popular “80 percent” figure is folklore, but even the measured number is large). For an operator, that translates to weeks of unglamorous work before anything intelligent happens.
Evaluation. You cannot ship what you cannot check. An eval, a repeatable test that scores whether the system is right often enough to trust, is the difference between a demo and a dependable tool, and building it is real work. Skipping it is the most expensive saving in the field: Gartner predicted at least 30 percent of generative AI projects would be abandoned after the proof of concept by the end of 2025, citing poor data quality and unclear business value among the causes (those failures are usually the unbuilt eval and the unprepared data, surfacing late).
Upkeep. This is the line operators miss most, because software is supposed to be a thing you finish. An AI system is not. Models get deprecated and replaced, prompts that worked drift as the model changes underneath them, and your own data shifts. So a meaningful slice of the build cost recurs every year as monitoring, re-evaluation, and maintenance. The dependable framing is that most of an AI system’s lifetime cost is ongoing, not upfront, which is exactly backwards from how a one-time software purchase feels.
There is a structural reason these stay expensive while tokens fall: tokens are a commodity priced by a competitive market racing to zero, and your integration, your data, and your evaluation are bespoke to you and priced by human time. Cheaper calls have not shrunk AI budgets in practice; they have expanded what people attempt, so total spend keeps rising even as the unit price collapses.
A worked budget (illustrative, on the cited prices)
Make it concrete with one representative case: a support-triage assistant that reads each incoming ticket, drafts a reply, and routes the hard ones to a human. The arithmetic below is illustrative, built on the live unit prices cited above, to show the shape of the budget, not to quote your quote.
Take the model line first. Anthropic’s own documentation works an example at about 3,700 tokens per support conversation on Haiku 4.5, landing near $37 per 10,000 tickets. An SMB handling 3,000 tickets a month is therefore spending on the order of $11 a month, call it $130 a year, on the actual intelligence. (That is the provider’s own figure, so treat it as a floor; real prompts with retrieved context and retries run higher.)
Now the parts that are not the model. Wiring the assistant into the help desk, giving it the knowledge base, building an eval against a few hundred real tickets, and getting it past pilot into daily use is a project measured in person-weeks, not dollars-per-million. Whatever that build costs, plan to spend a recurring fraction of it, a common rule of thumb is 15 to 25 percent annually, on monitoring and maintenance, indefinitely. The honest headline: the token line is a rounding error, the build is the cost, and the upkeep is the cost that comes back every year. (Gartner’s $5-to-$20-million range is for enterprise transformations, not an SMB triage bot; the structure scales down, the digits do not.)
When it pays back, and what to measure
Set expectations against evidence, not vendor decks. Deloitte’s late-2025 survey of 1,854 executives found most needed two to four years to reach satisfactory ROI on a typical use case; only 6 percent reported payback in under a year, and just 13 percent within twelve months, against the seven-to-twelve-month payback people expect from ordinary technology. Read that as a directional warning rather than a precise schedule, because it is self-reported and skewed toward larger firms, but the warning is real: AI investments pay back slower than the pitch implies, and a narrow, well-scoped SMB automation can beat that curve precisely because it is narrow.
To know where you land, measure the right things. Record the baseline before you build, the hours, the error rate, the cycle time of the process today, because you cannot prove a gain against a number you never wrote down (this is the same discipline that separates the firms getting return from the firms getting a bill: see who is actually getting return on AI). Then track the all-in cost, not the API cost: build, integration, and the annual upkeep, because the people who only watch the token meter are watching the cheap line. Most organizations cannot do this yet; CloudZero, a cost-management vendor, found only about half of companies strongly agree they can track their AI ROI at all. The ones who can are the ones who decided what payback would look like before they started.
The one-number rule
If you take a single habit from this, take this one. When a vendor or a quote puts the token or subscription cost in front of you, mentally multiply it by something like ten and treat that as your year-one budget, with the gap going to integration, data work, evaluation, and a year of upkeep. The exact multiple is yours to tune, and it shrinks as you build the second and third system on the same plumbing. But starting from “the model call is a small fraction of the real cost” will put you closer to the true number than any pricing page will, because the pricing page is advertising the one part of the project that was never going to break your budget.