Why agency prices have not followed the fall in AI costs

What cost $60 per million tokens in late 2021 now costs between $0.06 and $0.40 at equivalent performance, a reduction of 150 to 1,000 times. Meanwhile, a developer's average daily rate in France remains around €629. Someone is keeping that margin.

What cost $60 per million tokens in late 2021 now costs between $0.06 and $0.40 at equivalent performance (aisuperior.com/llm-token-cost). Yet an independent developer's average daily rate in France remains around €629 (rh-solutions.com, 2026 IT freelancer daily-rate guide). Someone is keeping that margin.

Inputs divided by 100, quotes at the same price

A CIO Online article documents the paradox among large accounts: developer productivity gains of up to 40%, a 30–40% reduction in timelines for standard features, but no reduction in bills (cio-online.com). Reasons offered: billed value never lies in raw code, hidden AI costs exist, and above all most buyers do not ask the question.

What an agency actually buys with your money

An agency account on dev.to summarises the client-side tension: a client requests a website ‘with AI’ and says ‘AI should make it possible for $10,000, right?’ for a project worth ten times as much (dev.to/akari_iku). Another leaves for a cheaper competitor ‘with AI’ and comes back disappointed. The feeling that something has changed in costs is correct; the bill has not followed.

‘AI makes mistakes’: true, but let's quantify it

The METR study measured experienced developers as 19% slower with AI while they believed they were faster (newsletter.getdx.com/p/metr-study-on-how-ai-affects-developer-productivity): human supervision really costs time, and that time justifies part of the price that has not fallen, an argument we take seriously rather than dismissing. But only part. One cited order of magnitude for AI production itself suggests tokens account for only 10–20% of an AI bill's total cost, the rest being infrastructure and supervision (bodic.fr, to cross-check for your own case). Even allowing for this safety margin, the gap between the scale of falling input costs and unchanged quotes remains huge.

When the cost reduction will reach you, and how

At Maestro, the reduction passes through directly: a complete product consumes a few dozen euros in artificial intelligence tokens, as we publish campaign after campaign, rather than a five-figure agency quote. This is not a marketing promise but the mechanical consequence of charging actual infrastructure costs rather than a daily rate inherited from before the reduction.

Pay for judgement, never again for typing

An agency's value is not keyboard typing but judgement: which architecture to choose, which compromise to accept, what to reject. That judgement retains its value, AI or not, and nobody should expect it to cost less tomorrow than today. What should fall is the price of typing itself, which models now perform at a negligible cost compared with a salary. As long as a quote mixes the two without distinguishing them, lower input costs appear nowhere on the final line. For more on the real cost of a project led by an agent team, the details are public.