AI prices dropped 97% since 2023. So why are AI bills 3x higher?
Something I keep running into with engineering and GTM teams: the per-token price of AI has fallen roughly 97% since early 2023. Models that cost hundreds of dollars per million tokens now run under $1. And yet enterprise AI bills are about three times higher than two years ago. Both things are true at once, and the "AI saves money" pitch never accounted for it.
Token spend per company is up 13x since January 2025. Goldman Sachs projects agentic AI could push token consumption up 24x by 2030. The price keeps falling. The bill keeps climbing.
Why falling prices don't mean falling bills
It's cloud storage in the early 2010s all over again. Cost per gigabyte cratered year after year, and AWS revenue still grew every quarter, because the amount people stored exploded. Price per unit dropped, total spend soared. AI billing is following the same arc, just faster. Once vendors moved from seat licenses to usage pricing, your bill stopped tracking headcount and started tracking consumption. And consumption is outrunning price cuts by a wide margin. (This seat-to-usage shift is reshaping SaaS pricing well beyond AI vendors, we dug into that here: B2B SaaS Pricing in 2026.)
Agentic workflows are the real multiplier
A single chat interaction is cheap, a few hundred tokens. Agentic workflows are a different beast. Reasoning models generate long thinking chains, agents call themselves in loops, context windows grow with each iteration. One agentic coding task can trigger 5 to 20 model calls and burn 10 to 50x the tokens of a chat exchange. Gartner pegs agentic workloads at 5 to 30x the compute of a standard chatbot call.
The uncomfortable part: the more autonomous and "efficient" your AI usage looks on the surface, the more expensive it quietly becomes underneath. Uber reportedly burned its entire 2026 AI coding budget in four months and had to cap spend at $1,500 per employee per month. If they didn't see it coming, most mid-market teams are flying blind.
The costs that never show up on the invoice
Model licensing feels like the main expense. It's typically around 20% of true total cost of ownership. Data preparation absorbs up to 45% of project effort in many organizations. Integration work, prompt drift, shadow AI usage, uncontrolled agentic overruns: none of it appears on a per-token invoice. It shows up months later as engineering hours and budget conversations nobody wants to have. (It's also why the build-vs-buy math surprises so many startups: agency vs. building in-house.)
The fix is operational, not strategic
Freezing budgets or rolling back AI is the wrong response to the wrong diagnosis. Three levers actually work:
Routing: not every task needs a frontier model. Sending simple tasks to smaller models cuts costs 30 to 85% without degrading quality on those tasks.
Caching: semantic caching serves stored results when a new prompt is functionally identical to a previous one. On high-repetition workflows it saves 30 to 90%, and almost nobody does it.
Visibility: if you can't trace which team, workflow, or product line drives spend, you're reacting to bills instead of shaping usage. This is what makes the other two levers possible.
The teams solving this aren't the ones spending less on AI. They're the ones spending smarter.
Curious what this community is seeing: is anyone actually tracking token spend per workflow? And if you've implemented routing or caching, what savings did you actually get versus the claimed ranges?
The metric I trust is cost per verified outcome, not cost per token or even per workflow. A run is only cheap if it ends with the expected artifact and checks; retries, abandoned branches, and false completion messages belong in the cost too. Hard ceilings prevent the worst bill, but outcome-level metering shows which agent loops are actually earning their spend.
The useful unit is rarely a token or even a model call. For DictaFlow, we look at the cost of a completed dictation workflow, because the expensive part can be retries, refinement, or long context the user never sees. Metering by workflow also makes it easier to spot where the extra spend is actually buying a better result, versus where an agent is just taking the scenic route.
Routing's the one lever here I can actually back up firsthand. Picked a smaller model on purpose for a narrow, well-scoped task instead of defaulting to whatever's newest, and it holds up fine, the savings are real when the task doesn't need frontier reasoning.
None of the three levers protect you from the case where usage stops behaving. A bug in a loop. Someone hammering an endpoint. You don't get a warning, you get a bill.
So there's a circuit breaker underneath everything else now, a hard daily spend ceiling that just returns an error once it's crossed, independent of any per-user limit. Doesn't route anything, doesn't cache anything, just guarantees the number can't go past what I picked no matter what the smarter layers miss.
Is anyone actually running a hard ceiling like that, or is everyone trusting routing and caching to catch it before the invoice does?
Great point. Falling model prices don’t automatically translate into lower AI costs because usage patterns are changing faster than efficiency gains. The shift from simple prompts to autonomous workflows completely changes the economics.
I think visibility will become one of the biggest challenges — companies need to understand not just how much they spend on AI, but where and why that usage creates value. Smarter routing, monitoring, and optimization will likely become essential parts of AI operations.
Falling price + rising bill = cloud storage in 2012 all over again. The real killer is agentic loops — one "efficient" agent can 20x a chat call and you don't see it till the invoice. Meter per-call, not per-session. Are you tracking spend per workflow, or just the monthly number?
Arleen, the cloud storage analogy is terrifyingly accurate.
As a solo founder running a fully automated content factory (managing 44 automated YouTube channels and AI blogs via n8n and Python), I felt that "agentic loop" token burn firsthand.
When I first set up my pipelines, I was throwing top-tier models at every single step—from deep reasoning to basic JSON formatting. The bill exploded overnight because the autonomous loops were just talking to each other endlessly.
The real game-changer for me was exactly your first point: Routing. I had to redesign my architecture to route the heavy lifting (script generation, logic) to advanced models, and strictly delegate the simple formatting/translation tasks to smaller models or just pure Python logic. Tracking token spend per n8n webhook node became mandatory for my survival.
Quick question regarding caching: Have you found semantic caching effective for dynamic content generation (like scripts or blogs), or does it mostly shine in static Q&A workflows?
Excellent breakdown of the reality we are all facing.
The real insight is that cheaper tokens don't mean cheaper bills — they mean more ambitious apps. When inference was expensive, you thought twice about every call. Now you throw 50 system prompts at a problem before finding the right one. The waste moves from compute to prompt engineering iteration cost.
I think the cloud storage comparison is spot on. Lower unit costs often encourage more usage instead of lower spending. I'm building an AI-powered business myself, and it's made me realize that monitoring usage patterns is just as important as choosing the right model.
I've tried almost all the best token-saving methods currently available on GitHub, and I must say that excellent projects like Headroom, Codegraph, and RKT are indeed helpful. My question is, how do I handle cross-session workflows when I have to run
clearwith a huge amount of historical information caching?The falling token price is only half the story. What ultimately matters is whether AI shifts from being a technology cost to an operational cost. Once usage becomes embedded in workflows, the biggest optimization isn't cheaper models—it's deciding which work deserves AI in the first place.
This comment was deleted 6 hours ago.