Inference Costs Fell Roughly 280x in 18 Months. Rebuild Your Business Case

Stanford's AI Index 2025 reported that the cost of achieving GPT-3.5-level performance on a standard benchmark dropped from approximately $20.00 per million tokens in late 2022 to around $0.07 by late 2024 — roughly a 280-fold reduction in about eighteen months. Hardware costs fell around 30% a year and energy efficiency improved around 40% a year over the same period.

This matters less for what it enables technically than for what it invalidates commercially. Every AI business case built on 2023 pricing assumptions is wrong by orders of magnitude. Use cases that were rejected as uneconomic — classifying every inbound email, enriching every lead record, summarising every support conversation, checking every document — are now rounding errors against a single hour of staff time at €31.22.

Recalculate honestly. Processing 10,000 documents at a few thousand tokens each is a few euro of inference. The cost of that project now sits almost entirely in integration, validation and change management. The model is the cheapest line item, and it is getting cheaper while your labour cost rises 3–4% a year.

The index also found that industry produced nearly 90% of notable AI models, and that the performance gap between the leading proprietary model and the best open-weight alternative narrowed sharply. For a buyer that means less lock-in risk than the market implies: designing so the model is a swappable component behind your own interface costs little and preserves the option to switch as prices continue to fall.

The strategic error is to spend the saving on more model calls. Falling inference cost does not make an ungoverned process valuable — it makes a bad output cheap to produce at volume. The constraint has moved decisively to data quality and evaluation: can you tell whether the output was right, and can you catch it when it is not?

So build the cheap thing, and spend the budget you were going to spend on tokens on the review loop instead. That is the allocation the cost curve now argues for.