Two facts about AI costs are both well documented and appear to contradict each other.
Prices have collapsed. When GPT-4 launched in March 2023 it cost $30 per million input tokens. By mid-2026, equal-or-better quality is available for under $0.50 — a fall of roughly 95%. Epoch AI's tracking puts the decline at somewhere between 9× and 900× per year depending on which capability milestone you hold constant. Aggregate API pricing fell about 80% between early 2025 and early 2026 alone.
Bills have gone up. Enterprise spend on LLM APIs passed $8.4 billion in 2025 and is on track to double again.
Falling unit price, rising total spend. If that seems paradoxical, it is only because the industry keeps discussing cost in the units vendors publish — dollars per million tokens — rather than the units that produce a bill.
Where the tokens went
The reconciliation is not subtle. Gartner's March 2026 analysis puts agentic workloads at five to thirty times more tokens per task than a standard chatbot turn. An unoptimised production agent has been costed at $10–100 per session.
So the shape of consumption changed at roughly the same rate as price, in the opposite direction.
| Variable | Direction | Rough magnitude |
|---|---|---|
| Price per million tokens GPT-4-class | down | ~95% |
| Tokens per task agentic vs. chat | up | 5–30× |
| Enterprise LLM spend | up | ~2× and rising |
A 95% price cut is a 20× saving. A 30× consumption increase more than eats it. Both trends are real; the second is simply larger and less discussed.
Cheaper tokens didn't reduce spending. They made a more expensive way of working affordable enough to adopt.
This is a well-known pattern outside software — it has a name, Jevons paradox: efficiency gains that lower the cost of using a resource tend to increase total consumption of it rather than reduce it. Cheap inference did not make existing workloads cheaper so much as it made agentic workloads viable, and agentic workloads consume on a different scale entirely.
Why the price sheet stopped being useful
The practical consequence is that reasoning about AI cost from published rates has become close to useless for anyone running agents.
A price sheet gives you four numbers per model — input, output, cache write, cache read — and your actual bill depends on a mix that varies enormously by workload. On my own agentic workload, cache reads were 97.94% of tokens consumed, so the effective blended rate was $0.65 per million against a $5 list input price. Someone doing short interactive chat turns would pay something close to list. Same models, same published rates, roughly an eightfold difference in effective cost.
Which means the sentence "the API costs $5 per million tokens" carries almost no information about what you will pay.
What this changes about optimisation
If unit price is falling faster than you can negotiate it, and consumption is rising faster than price is falling, then the leverage is entirely on the consumption side.
Where the remaining leverage is
- Context size per call — on agentic workloads this dominates everything else. What you keep in the window is the bill.
- Cache hit rate — cached reads bill at a tenth of input. A cache entry pays for itself on the second read, and a silent cache regression can multiply a bill without raising an error.
- How many calls a task takes — agent design decides this long before any model is chosen
- Not: waiting for prices to fall further. They will, and it won't help — that is exactly what the last two years demonstrated.
- Not: model shopping on headline rate alone. Routing 7% of my calls to cheaper models moved 0.89% of spend.
The uncomfortable version
There is a reading of this that vendors are unlikely to volunteer. Falling per-token prices are presented as a straightforward win for buyers, and at the unit level they are. At the invoice level, the last two years suggest that price cuts have been substantially recaptured by consumption growth — which is a fine outcome if the additional consumption is producing proportionally more value, and an expensive one if it is not.
Whether it is producing value is a separate question, and the available evidence there is not encouraging: MIT's study of 300 enterprise AI deployments found 95% delivered no measurable P&L impact. Cheaper tokens, more of them, and mostly nothing at the bottom of the page.
Forecasts point to another 90% cost reduction by 2030. On the evidence of the last two years, the right response to that is not to plan for smaller bills.
Method & sources
This piece is analysis of published figures, not original measurement. That distinction matters here: unlike the research on this site drawn from my own logs and invoices, every number below comes from someone else's work and carries their methodology and limitations.
Price-decline figures are from Epoch AI's inference price tracking and reporting on GPT-4-class pricing between 2023 and mid-2026. The agentic token multiple (5–30×) is Gartner, March 2026. Enterprise spend figures and the $10–100 per unoptimised agent session are from Zylos Research. The 95% pilot-failure figure is from MIT's study of 52 executive interviews, 153 leader surveys and 300 public deployments.
Aggregate figures like these vary by source and definition — "GPT-4-class" in particular is a judgement about capability equivalence, not a measurement. Treat the direction and rough magnitude as sound and any single decimal place as approximate.
The one first-party figure is the 97.94% cache-read share, from my own 34-day dataset. No commercial relationship with any vendor named, and no affiliate arrangement on anything in this piece.