Full narration of the briefing. Click a section heading to jump the player to that moment.
Google Cloud made two changes this year that moved risk toward the customer while presenting as flexibility. On June sixteenth the default scope of resource based committed use discounts changed from the project that bought them to the whole billing account. On August twenty sixth Gemini Enterprise gained three new ways to pay: pay as you go, Flexible Savings Plans and deferred execution. I am Claire, Tom is with me, and this briefing is how the new levers work, what they do to your chargeback model, the price step already in the calendar for January, and how to size any commitment on evidence rather than a sales forecast.
Agent workloads broke per seat pricing: one user request can trigger a chain of model calls, reasoning steps and retries that no seat count predicts. Google's answer is to unbundle the price. Pay as you go removes the base subscription and bills tokens and compute at standard rates. Flexible Savings Plans are spend commitments at ten percent off for one year and twenty for three, counting against an existing enterprise agreement.
Deferred execution marks eligible agent jobs for off peak processing at up to half price. For comparison, Compute Engine Flex CUDs run about twenty eight percent, and resource based CUDs thirty seven percent and above.
Look at the mechanics before the discount. Commit two hundred thousand dollars a month and use one hundred fifty thousand: the bill is two hundred thousand. Use two hundred fifty thousand: the extra fifty is billed at on demand rates with no discount. The plan protects Google's floor in both directions.
Three years is a bet on the market, not the workload. A three year Gemini commitment assumes the workload, the model roadmap and Google's pricing structure all stay favorable while model versions turn over every few months. Break even on a one year plan is roughly ninety percent utilization. On three years, the forecast horizon makes the number meaningless.
Now the quieter change. Resource based CUDs on Compute Engine used to cover usage in the project that bought them. From June sixteenth the default is billing account scope with sharing enabled. New accounts got it automatically, and so did existing accounts without active commitments.
A commitment bought by one business unit now covers eligible usage in any project on the account, allocated by proportional attribution unless you configure prioritized attribution. Nothing about that costs more in total. Everything about it changes who appears to pay: the unit that bought the commitment sees a higher effective rate, other units see savings they never bought, and the chargeback model finance built no longer describes the bill. Nobody was told.
There is a published price step in the calendar, and the account team's sizing model may not include it. Gemini three point six and three point seven Flash launched at introductory rates through December thirty first: seventy five cents per million input tokens and three dollars seventy five for output. From January first, 2027 the standard rates are one dollar fifty and seven fifty. Double.
A savings plan sized on the 2026 token price covers half the volume in 2027. A plan sized on the 2027 price is a one hundred percent overcommit in 2026. Size on the post step, post optimization run rate, and treat the Q4 2026 bill as the worst possible baseline.
Here is the illustration. An enterprise with a growing Gemini footprint is offered a three year plan in the last quarter of 2026: two hundred fifty thousand dollars a month, projected flat, twenty percent off, nine million committed. Year one holds. Then caching and routing cut tokens per outcome, one agent program is cancelled, and list spend settles at one hundred fifty thousand a month.
Years two and three spend three point six million at list against six million of commitment: two point four million paid for nothing, no rollover. Total, nine million paid against six point seven million of usage. The commitment cost two point three million more than no commitment.
The move from this briefing. Establish the baseline on pay as you go for a full quarter and export usage by model, agent and project. Reprice that baseline for January first. Route evaluations, indexing and document processing to deferred execution first: half off with no term beats both savings plans.
Then take one year, not three, sized at seventy to eighty percent of the repriced run rate. Audit every billing account for the June switch, set attribution deliberately with finance, and fix in writing that plan spend burns down the enterprise agreement at full value. The full research note is free to download under this video, at redresscompliance dot com slash newsletter slash september.
Redress Compliance works on contingency: our fee is 25 percent of what we save you. Nothing saved, nothing paid. Independent, buyer side only, never vendor funded. Want Redress to contact you? Reach out and we respond the same day.
Talk to a Google Cloud negotiator