Small business owners budget for AI the way they budget for software: a flat monthly number that stays put until someone calls to renegotiate.

That’s the wrong model. AI pricing moves on its own, in both directions, and it doesn’t ask.

What is about to happen

Google’s flagship reasoning model costs $2.00 per million input tokens and $10.00 per million output tokens. Those prices are listed as valid “through November 30, 2026”. On December 1 they become $4.00 and $20.00.

That’s a doubling, published in advance, on the provider’s own pricing page, and it applies automatically to anyone whose account is already running. No email asking you to agree. No renewal date to push back against.

Plenty of owners haven’t noticed this one yet. It arrives whether or not you did.

The other way it bites you

Price is only half of it. The model itself has an expiry date.

Anthropic states plainly that it retires older models “to ensure capacity for new model releases”, and that requests to a retired model will fail. It commits to at least 60 days’ notice by email for publicly released models. Two of its older Opus models are already listed as retired everywhere except two cloud platforms.

Sixty days sounds generous until you work out what it means for a business. Your assistant worked fine on Friday. On Monday the integration returns an error, and “reboot it” isn’t a fix. Someone has to re-test everything against a different model, because the new one answers differently.

What this means for a plan

Two rules follow directly, and both are cheap to adopt now and expensive to retrofit.

Put a ceiling on anything usage-priced. Every provider lets you set a spending cap on an API key or workspace. Set it. A tool that can spend an unlimited amount isn’t a subscription with a variable bill, it’s a bill with no ceiling. If a tool won’t let you cap it, keep it off anything automatic.

Write down which model you’re actually using. Not the product name. The model. If you can’t find it, you have an unnamed dependency in your business, and when it changes you’ll find out from an error message rather than a plan.

What to do before December 1

Pick your three most-used AI tools and answer three questions about each:

  1. Which model does it call?
  2. Is it billed per use, or flat?
  3. If that model disappeared in 60 days, who would notice and how long would the fix take?

Two of the three will be flat-rate products where the provider absorbs the change. That’s a fine answer. The third one is usually something you wired up yourself a year ago, and it’s the one that needs a name and a cap before it needs a fire alarm.

The tools that survive this exercise are the ones you can put in a budget. The ones that don’t are the reason AI line items keep coming in over.