Are AI Prices Actually Falling?
AI pricing guide · updated 2026-08-07
The received wisdom is that AI inference gets cheaper every month. Our own change log tells a more awkward story: cuts do outnumber increases, and the average move is still close to zero. Prices churn in both directions, and a model you already depend on is about as likely to have gone up as down.
What the log shows
As of this update, the log held 278 changes to tracked fields since early May 2026: 166 cuts and 112 increases, accruing at roughly three a day. The counts grow daily; the balance has stayed close to three cuts for every two increases. The average move is about minus 1 per cent, which is essentially flat.
Underneath that average, individual repricing is substantial: 42 of those changes moved a rate by 20 per cent or more in one direction or the other. So the headline calm conceals real volatility at the level of specific models.
The live figures behind this sit on our price-change page, which regenerates as the daily refresh finds new moves.
Why "prices always fall" feels true
Two things create the impression. New models frequently launch cheaper than the models they replace, and the cheap end of the market keeps getting cheaper, so the floor drops even when established rates hold still.
Both are real, and neither means your bill falls. A cheaper new model only helps if you migrate to it, and migration means re-testing prompts, re-running evaluations and accepting different behaviour. The floor dropping is an opportunity, not a discount.
Why prices go up
Introductory pricing expiring is the most common cause. Claude Sonnet 5 launched at $2.00 and $10.00 per million tokens with those rates stated to run until the end of August 2026, after which they move to $3.00 and $15.00. That is a 50 per cent increase arriving on a scheduled date for anyone who priced against the launch figure.
Providers also reclassify models between tiers, change how long-context requests are billed, and withdraw promotional rates. None of these are announced loudly, and none of them show up until an invoice does.
What to do about it
Re-check the rate for any model you depend on before treating a months-old figure as current, and note the expiry date whenever you adopt something on introductory pricing.
Watch the floor as well as your own rate. The saving available from moving down a tier is usually far larger than any change to the model you are on, because the catalogue spans about 1,600 times from cheapest to dearest.
And design for substitution. A system where the model is a configuration value rather than an assumption can take advantage of a cheaper option in an afternoon, which is what turns a falling floor into an actual reduction.
Frequently asked questions
Are AI API prices going down?
Less consistently than people assume. Across the 278 changes logged between early May and early August 2026 there were 166 cuts and 112 increases, and the average move was about minus 1 per cent, which is close to flat. The price-change page carries the current figures. New models often launch cheaper than their predecessors, but established rates move in both directions.
Why would an AI model get more expensive?
Introductory pricing expiring is the most common reason: Claude Sonnet 5, for instance, launched at $2.00 and $10.00 per million tokens with those rates running to the end of August 2026, then moving to $3.00 and $15.00. Providers also reclassify models between tiers, change long-context billing and withdraw promotional rates.
How can I benefit from falling AI prices?
Treat the model as a configuration value rather than a fixed assumption, so switching is an afternoon of work rather than a project. The largest savings come from moving down a tier when a cheaper model clears your quality bar, since the catalogue spans roughly 1,600 times from cheapest to dearest, far more than any individual rate change.
See it in the data
Related guides
Terms used in this guide
Published by Tokenando. Last updated 2026-08-07. Figures in this guide are computed from our own pricing index and dated where they can move; see the methodology and corrections policy.