By Brad Ferris · 23 May 2026
The most commercially important announcement at Google I/O this week was not a model. It was a price.
Google launched Gemini 3.5 Flash at roughly one-third the price of existing tier-one models and cut its top AI subscription by around 20 per cent to US$200 a month, while adding a cheaper US$100 tier. Alongside the pricing, Sundar Pichai disclosed the scale this is riding on: Google is now processing 3.2 quadrillion tokens a month, roughly seven times what it handled a year ago, with the Gemini app past 900 million monthly active users.
Cheaper and better at the same time is not how enterprise software has historically behaved. It is how AI has behaved for eighteen straight months, and this week made the trend impossible to ignore.
Two other data points from the same week complete the picture. Simon Willison, one of the most careful independent voices in this field, summarised the last six months of models and noted that the "best model" crown has changed hands five times across three different providers since November, while open-weight models you can run on a laptop have improved far past expectations. And Anthropic agreed to pay roughly US$1.25 billion a month for compute capacity through 2029, renting it from xAI, a direct competitor. Even the labs cannot own enough of the layer beneath them, so they rent it and stay flexible.
Take the two facts as one. Prices are falling fast, and no vendor holds the capability lead for more than a few months. For a business buying AI, that combination points to a clear posture.
Business cases you shelved deserve a rerun. If you costed an AI project in 2025 and parked it as too expensive, the inputs are stale. A workload priced against last year's per-token rates can look completely different against a model a third of the price. This matters most for the high-volume, low-margin ideas: document processing, call summarisation, catalogue work. Things that failed the maths at old prices often clear it at new ones. Rerun the numbers before you rerun the debate.
Long lock-ins are the wrong trade. Signing a multi-year, single-vendor AI commitment right now means paying today's prices for a capability that will be cheaper and probably beaten by a rival within two quarters. Where a vendor insists on term, insist on price-review clauses and the right to substitute models. Willison's five-crowns-in-six-months observation is the negotiating evidence.
Design for switching from day one. The practical version of vendor flexibility is architectural, not contractual. Keep your prompts, workflows and evaluation tests documented outside any one vendor's platform, so that moving from one model to another is a re-test rather than a rebuild. Businesses that did this in the past year have quietly upgraded three times for free. Businesses that hard-wired one vendor's quirks into their processes have not.
Expect your total spend to rise anyway. This is the honest footnote. Unit prices are collapsing, but usage grows faster. Google's seven-fold token growth is the pattern in miniature: as AI gets cheaper, you use it for more things, and the bill reflects volume, not rates. Budget for adoption growth, and put someone in charge of watching usage the way you watch mobile or cloud spend.
Falling prices did not stop serious buyers from committing this week. KPMG announced it is rolling Claude out to its entire 276,000-person global workforce. The lesson is not that big firms ignore lock-in risk. It is that they have decided waiting costs more than switching later will. The price curve rewards businesses that start now with flexible arrangements, not businesses that wait for the curve to flatten. It shows no sign of flattening.
For an Australian mid-market operator running tight margins, the takeaway fits on a sticky note. Buy short, design portable, measure usage every month, and rerun last year's rejected business cases. The intelligence keeps getting cheaper. The advantage goes to whoever has built the habits to keep re-buying it well.
Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.