There's a version of the AI story that goes like this: prices collapse, access democratizes, small operators can finally afford what enterprises were paying millions for, and everyone benefits equally. That version is half right.
Prices have collapsed. The access is real. But the bill shock is also real — and it's showing up exactly where founders are building the most. The companies most actively deploying AI workflows are, in many cases, spending more than they were when AI was ten times as expensive. That's not a glitch. It's a well-documented economic pattern, and understanding it is the difference between AI being a cost-controlled competitive advantage and a budget line that erodes the margin you were trying to protect.
The Real Numbers Behind the AI Price War
Let's start with what's actually happened to LLM pricing in 2026, because the scale of the drop is legitimately remarkable.
On July 30, OpenAI cut GPT-5.6 Luna — its most widely deployed commercial model — by 80%. That same week, Google's Gemini Flash was already sitting at $0.075 per million input tokens, down from over $7 per million on comparable capability just three years ago. Anthropic cut Claude Sonnet pricing by over 60% within months of launching updated versions. The floor for mainstream LLM APIs now sits near $0.20 per million input tokens. DeepSeek and smaller providers are running even lower.
To put that in concrete terms: in 2023, running a workflow that processed one million tokens — roughly 750,000 words, or the full text of about 10 business novels — through a frontier model cost somewhere between $20 and $60. Today the same volume through Gemini Flash costs about 7 cents. If you're using a mid-tier model for most tasks, a workflow that would have cost $500 per month to run in 2023 now costs $10 to $20.
CloudZero, which tracks AI infrastructure spend across hundreds of organizations, calls this the most rapid unit cost collapse in the history of commercial software infrastructure. Faster than cloud storage. Faster than compute. Faster than bandwidth. The per-token price has fallen roughly 90% in the last 18 months alone.
This is genuinely good news. It means workflows that were economically marginal a year ago — things like processing every inbound lead through an enrichment and qualification sequence, or running a nightly summary of all customer support tickets against a CRM — are now trivially cheap to operate. Capabilities that lived behind enterprise contracts are now accessible to a five-person business paying API costs directly.
So why is aggregate AI spending going up?
An Economist Named Jevons Predicted This in 1865
William Stanley Jevons was studying coal in Victorian England when he noticed something counterintuitive: as steam engines became more efficient — consuming less coal to produce the same amount of work — total coal consumption in Britain went up, not down. The efficiency gain didn't reduce demand. It unlocked new uses, made coal-powered processes economically viable in places they hadn't been before, and drove total consumption higher even as per-unit efficiency improved.
Jevons published this as a formal economic observation in 1865. Economists call it the Jevons Paradox: when a resource becomes more efficient or cheaper to use, consumption tends to increase in ways that offset, and often exceed, the savings from the efficiency gain.
It applies to AI in 2026 with almost textbook precision.
Fortune reported in June that top economists are explicitly warning about Jevons dynamics in AI spending. The observation: tokens are 48 times cheaper than they were, but organizations are consuming 100 to 1,000 times more of them. Net spend per developer and per organization is rising two to twenty times despite the price collapse. This isn't a forecast — it's a pattern being measured in real billing data across major AI platforms right now.
The mechanism is straightforward. When something gets cheap enough to use freely, people use it for things they never would have attempted when it was expensive. A founder who was spending $200 a month on AI and carefully rationing model calls starts running AI across every lead, every email thread, every support ticket, every piece of content in the pipeline. The number of API calls doesn't stay the same. It multiplies — often by an order of magnitude. And while the per-call cost dropped 90%, the number of calls went up 1,000%.
The result is a higher bill than the one you were managing before.
What This Looks Like When It Hits a Growing Business
The pattern has a recognizable shape. A founder starts deploying workflows in earnest — usually because the cost justification finally feels obvious, which it now does. Three months in, the AI stack is doing something genuinely useful: qualifying leads, drafting proposals, summarizing calls, monitoring competitor updates. The monthly AI cost at that point is probably $150 to $300, down from a theoretical $1,500 if they'd built the same thing two years ago. So far so good.
Then the team starts using it for more things. The marketing person realizes they can use the same Claude API connection to draft social posts. The ops person builds a new workflow to process vendor invoices. The founder adds a research agent that runs weekly competitive intelligence. Each individual addition is cheap. Collectively, the consumption grows 5x in 60 days.
This isn't recklessness. It's the rational response to cheap, capable AI. The problem is that most founders don't have usage monitoring in place until after they've been surprised by a bill. And because the price per token is legitimately low, the absolute dollar amount of the surprise is often modest — $400 versus an expected $80 — but the pattern, if left unaddressed, scales with the business.
CloudZero's 2026 survey of 260 finance leaders found 55% ran over their AI budget, with 32% running more than 20% over. That overspend isn't happening because AI got expensive. It's happening because usage grew faster than the mental model did.
Anthropic, OpenAI, and Google all moved in April 2026 to impose tighter usage caps and clearer consumption dashboards on their commercial plans — a direct response to the bill shock complaints they were fielding from customers who had expanded into workflows without corresponding budget adjustments. The platforms saw it coming. The customers mostly didn't.
The Governance Framework That Prevents the Surprise
The fix is less technical than it sounds. It's essentially three things: a usage inventory, a monthly budget per workflow, and a monitoring trigger. None of this requires engineering. It requires treating AI usage the way you treat any other operational cost — with visibility and a defined ceiling.
Build a usage inventory before you scale. List every workflow or integration that makes API calls. For each one, estimate the call volume per day and the average token count per call. Multiply by 30 and by the model's cost per token. This takes about 20 minutes in a spreadsheet and gives you a baseline cost forecast for your current stack. Most founders who do this for the first time discover that 80% of their AI cost comes from two or three workflows — almost always the highest-volume ones, not the most sophisticated ones. A lead enrichment workflow running 300 times a day costs more than a weekly competitive intelligence report, even if the report uses a more powerful model.
Set a per-workflow cost ceiling and monitor it weekly. Every major AI provider now offers usage dashboards with the ability to set alerts. OpenAI's usage dashboard lets you set spending limits per API key. Anthropic's console has per-workspace monitoring. Google Cloud's AI usage tracking integrates with billing alerts. Use them. Set an alert at 80% of your monthly AI budget so you see the warning before the bill arrives. If you're building in Make.com or n8n, add a step that logs token usage per run to a Google Sheet or Airtable — it's a two-minute addition and gives you a running total you can check without logging into each provider separately.
Match model choice to volume, not just task complexity. This is where the Jevons dynamic gets correctable. The workflows generating the most token volume are almost never the ones requiring the most sophisticated reasoning. Lead classification, document formatting, email routing, ticket summarization — these are high-volume, low-complexity tasks that run fine on Gemini Flash ($0.075/M), GPT-5.6 Luna ($0.20/M after the July cuts), or Claude Haiku 4.5. Running these on frontier-tier models (Claude Opus, GPT-5.6 Terra, Gemini Ultra) because the reasoning capability occasionally helps with edge cases is the single most common cause of AI budget overruns in the businesses we work with. The edge cases don't justify the per-call premium at the volumes these workflows run. Route the edge cases to escalation — a human review queue or a higher-tier model call — and run the volume on the cheapest model that handles the core case correctly.
Audit quarter-on-quarter, not just month-to-month. AI usage doesn't grow in a straight line. It often plateaus, then jumps as a new workflow goes live, plateaus again, then jumps again. A month-over-month comparison misses that pattern. Pull your quarterly usage data every 90 days and compare it against what you expected when you launched each workflow. Anything that's running 3x the projected volume is either more valuable than you estimated — in which case the cost is probably justified and you should formalize it — or it has a runaway call pattern (a retry loop, a webhook misconfiguration, an agent calling itself) that you need to diagnose before it compounds.
The Opportunity Founders Are Missing While Worrying About the Bill
Here is the part that gets lost when the conversation turns to overspend: the same Jevons dynamic that creates billing surprises is also creating a legitimate competitive window that founders who set up governance correctly can exploit.
Eighteen months ago, the workflows that are now running for $30 to $100 per month — lead enrichment, proposal drafting, client reporting, competitive monitoring, invoice processing — were either too expensive to build at volume or required enterprise contracts with Salesforce, HubSpot, or Zendesk to access bundled AI features at scale. The price collapse has made founder-built equivalents economically viable in a way they genuinely weren't before.
The businesses getting ahead right now aren't the ones spending the most on AI. They're the ones who identified five to eight high-frequency, high-value workflows, built specialized agents scoped tightly to each one, priced the stack correctly using tiered models, and set up governance before they scaled. Their AI infrastructure costs $150 to $400 per month and covers what a $60,000-per-year hire would have covered two years ago. The margin improvement is measurable and defensible.
The businesses getting hurt are the ones who heard "AI is cheap now" and treated that as a signal to remove cost scrutiny entirely. Cheap per token doesn't mean the bill doesn't matter. It means the governance is easier to build — because the numbers are smaller and the tools are better — but it still has to be built.
Jevons' insight from 1865 is still right: efficiency gains unlock new consumption rather than reducing it. The question is whether that new consumption is governed, measured, and pointed at things that generate return — or whether it's sprawl that looks like progress and shows up as a budget problem at the end of the quarter.
If you want a clear-eyed look at your current AI stack — what it costs, what it's generating, and where the governance gaps are — bring us the details. We'll audit it honestly, tell you what's working, and flag anything that's headed toward a billing surprise before it arrives. We'd rather tell you to cut a workflow than let you scale one that's burning money. Start a conversation.
Related: Salesforce Built 7 Specialized AI Agents at Dreamforce. The Lesson Applies at Founder Scale. | AI Costs Dropped 280x. Your Bill Didn't. Here's Why.