On August 1st, Google Cloud quietly flipped a switch that a lot of small and mid-size businesses had been waiting on: Gemini Enterprise's pay-as-you-go tier went generally available. Instead of committing to a block of per-seat licenses before you know whether an AI rollout will actually work, you can now pay for the agent usage you generate, with built-in spend caps and usage monitoring baked into the console. A few days later, Google shipped expanded tracing for Gemini Enterprise's data connectors, adding visibility into exactly which tool calls and connector invocations are burning through that spend.
On paper, this is exactly what smaller companies have been asking for. Per-seat AI licensing punishes you for piloting — you either buy ten seats you don't need to test one workflow, or you don't test it at all. Pay-as-you-go removes that upfront commitment. But treat this as the moment your AI cost problem quietly resolves itself, and you'll be disappointed. The pricing model was never what was pushing budgets over the edge. What's pushing budgets over the edge is a lack of anyone watching the meter.
The Number That Should Worry You More Than the Pricing Change
Here's the part of this story that doesn't make it into the release notes: blended token prices across the major model providers fell roughly 67% year-over-year, from about $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026. By every normal rule of economics, AI spend at most businesses should be falling too, or at least growing slower than usage. Instead, recent industry research puts the share of companies exceeding their AI budget projections at 73% — nearly three out of four.
The actual problem
Tokens got six times cheaper in a year and most companies still blew their budget. That's not a pricing problem. That's a usage-governance problem — nobody scoped what "normal" usage looks like, nobody's watching for runaway loops or over-eager retries, and nobody set an alert before the invoice arrived. A cheaper meter doesn't help if no one's reading it.
Why Pay-As-You-Go Doesn't Fix This by Itself
Spend caps and usage dashboards are genuinely useful additions to Gemini Enterprise, and Google deserves credit for shipping them alongside the pricing change rather than as an afterthought. But a spend cap that fires after you've already burned through the month's budget is a smoke alarm, not a sprinkler system. And a usage dashboard is only useful to someone who checks it, understands what a normal week of agent activity looks like, and knows which spike is a real business surge versus a mis-configured agent stuck in a retry loop.
Most SMBs don't have a FinOps team, or even a single person whose job includes watching AI spend the way a controller watches cash flow. The tooling improved. The staffing gap that makes the tooling useful didn't.
- Pay-as-you-go removes the up-front seat commitment — good for piloting, but says nothing about whether a live agent is calling a tool three times when once would do
- Spend caps stop the bleeding after the fact — they don't catch a badly scoped prompt burning 5x the tokens it needs before the cap trips
- Usage dashboards show you what happened — they don't tell a non-technical business owner which line item is normal growth and which is a bug
- Connector tracing shows which tool got called — someone still has to know what the right call volume should have been to spot the anomaly
What a Genuinely Governed AI Rollout Looks Like
The businesses that keep AI costs predictable in 2026 aren't the ones that found the cheapest per-token rate — token prices are converging across providers anyway. They're the ones that treated cost governance as part of the build, not an afterthought bolted on after the first surprise invoice. In practice that means a handful of concrete things, not a vague commitment to "monitor usage":
- 1Smart model routing — sending simple, high-volume tasks to a cheaper, faster model and reserving the expensive frontier model for the calls that actually need its reasoning depth
- 2Response caching for repeated or near-duplicate queries, so the same customer question asked a hundred times a week doesn't generate a hundred full inference calls
- 3Prompt efficiency passes — trimming bloated system prompts and unnecessary context that quietly inflate every single call's token count
- 4A pre-agreed usage ceiling per workflow, set before launch, with an alert that fires at 70% of that ceiling — not a hard cap that just stops the agent mid-task
- 5A monthly line-item review of AI spend by workflow, the same way you'd review any other recurring vendor cost, so a slow creep gets caught in week two, not month four
None of this requires an in-house FinOps hire. It requires someone who scoped the AI system to include cost controls from the start, rather than shipping a working agent and hoping the bill stays reasonable.
How Wizeb Approaches This Differently
This is precisely the gap Wizeb's AI cost optimization work is built to close, and it's the reason we build it into every agent and automation project rather than selling it as a separate add-on after the fact. When we scope a custom AI agent, model routing, caching, and prompt efficiency aren't a later phase — they're part of the initial architecture, because a system that's expensive to run at 10 customers gets prohibitively expensive at 500. We've seen client AI costs drop 40–70% purely from routing and caching changes that didn't touch the customer-facing behavior of the agent at all.
We also don't leave clients holding a usage dashboard and hoping they know what to do with it. Every agent we deploy ships with a usage ceiling agreed before launch and a monthly review of what it actually cost to run against what value it delivered — the same discipline you'd want from any vendor relationship, applied to a technology that's genuinely new enough that most businesses haven't built that habit yet.
A realistic scenario
A 40-person professional services firm piloted an internal document-summarization agent on a per-seat enterprise AI plan and watched the bill triple in six weeks — nobody had scoped what "normal" usage looked like, and a handful of staff were re-running full documents through the agent multiple times rather than trusting the first output. Wizeb rebuilt the workflow with a cheaper routing model for first-pass summaries, cached repeated document sections, and added a usage ceiling with a weekly Slack digest showing spend by team. Total AI spend dropped 58% in the first month, with output quality unchanged — the fix was governance, not a different model.
A Three-Question Budget Check Before You Adopt Pay-As-You-Go
If your business is considering Gemini Enterprise's new pricing tier, or any pay-as-you-go AI product, ask three questions before you flip the switch:
- 1Who on your team will actually look at the usage dashboard weekly, and do they know what a normal week of activity looks like well enough to spot an anomaly?
- 2Is there a usage ceiling set per workflow before launch, with an alert before you hit it — or only a hard cap that stops your agent mid-task after the damage is already done?
- 3Has anyone reviewed the actual prompts and tool-call patterns for waste — duplicate calls, bloated context, an expensive model doing a cheap model's job — or did the system ship as soon as it worked?
If you can't answer all three with confidence, the pricing model you choose is the least important decision you're about to make. Token prices will keep falling — that trend isn't going to reverse. Whether your AI spend falls along with it depends entirely on whether anyone built governance into the system, not on which billing tier you picked. If you want a free review of what your current or planned AI spend actually looks like against what it should look like, that's a conversation worth having — start at wizeb.com/services/token-optimization.