Is The AI You Pay For The AI You Are Getting?
In July 2026 Anthropic was caught quietly routing paying subscribers onto a weaker model while continuing to bill them the full rate. The routing happened under an internal tag. When a user went digging in the logs and found it, a team member's reply included the line that they had not expected anyone to look.
I am not writing this to pile on a company. I use their tools. I am writing it because it exposes a structural problem that every Santa Clarita business buying AI needs to understand, and it has nothing to do with that particular vendor.
The thing you are buying is invisible
When a mechanic replaces a part, you can look at the part. When a contractor pours a slab, you can measure it. When you buy AI, you get text on a screen, and text on a screen looks the same whether it came from the best model available or a cheaper one running at a fraction of the cost.
You have no native way to tell the difference. That is not a conspiracy, it is just the shape of the product. And any product where the buyer cannot verify what they received creates a standing temptation on the seller's side, because serving you something cheaper is pure margin and almost nobody checks.
Why this hits small businesses hardest
An enterprise buyer has a contract with named models, committed capacity, and a procurement person whose job includes noticing this. You have a credit card on file and a monthly charge.
You are also the buyer least likely to have a baseline. If output quality degrades ten percent over six weeks, an enterprise sees it in a dashboard. You experience it as a vague sense that the tool is not as sharp as it used to be, and you assume you are imagining it. Usually you conclude you are the problem.
How to actually verify what you are getting
Build a small fixed test set and run it on a schedule. This is the whole answer and it takes about an hour to set up once.
Write five to ten prompts that represent the work you actually use AI for, and that have answers you can judge. Not trivia. Real tasks: summarize this kind of document, draft this kind of reply, pull the key terms out of this contract. Run them, save the outputs with the date, and put a recurring reminder on your calendar to run them again monthly.
You are not trying to score the model. You are building a baseline so that drift becomes visible. Quality changes that are invisible in daily use are obvious when you put January's output next to August's on the same prompt.
The version that is harder to catch
Most Santa Clarita businesses are not buying AI directly. They are getting it inside something else, bundled into a CRM, a scheduler, a marketing platform.
That is a harder problem, because the vendor can swap the underlying model whenever their costs change and has no obligation to tell you. Your only signal is that the summaries got worse. Run the same fixed test set against those embedded features too, and when quality moves, ask the vendor directly what changed. The quality of the answer to that question tells you a great deal about whether to keep paying them.
What to put in the agreement
Ask for the specific model or a committed capability floor in writing. Ask for notice before material changes. Ask for the right to leave without penalty if quality drops below the floor.
Vendors who are dealing straight will not fight you on these. Language about continuous improvement and evolving our model mix sounds reassuring and commits them to nothing at all. That distinction is worth reading for, because it is the only leverage you get before you have a problem rather than after.
The underlying rule
Assume you cannot see what you bought, then build the one small check that lets you see it.
This is not cynicism about AI vendors. It is the same discipline you would apply to any supplier whose product you cannot inspect on delivery. The tools are genuinely useful. Verifying what you receive is what keeps them useful.
Common questions
What happened in July 2026?
Anthropic was found to be quietly routing some paying subscribers onto a weaker model under an internal tag while still charging the full subscription rate. After the backlash the company extended a free usage grace period.
Does this mean AI vendors cannot be trusted?
It means the thing you are buying is invisible, so trust has to be verified rather than assumed. You cannot see which model answered you the way you can see which part a mechanic installed.
How can a small business tell which model it is actually getting?
Keep a small fixed set of test prompts with known good answers, run them monthly, and save the outputs. Quality drift shows up in that comparison long before you would notice it in daily use.
Does this affect AI features built into other software?
Yes, and more so. When AI is embedded in a CRM or a marketing tool, the vendor can change the underlying model without telling you, and your only signal is that output quality shifted.
What should be in a contract with an AI vendor?
The specific model or a committed capability floor, notice before material changes, and the right to leave without penalty if quality drops. Vague language about continuous improvement gives you nothing.
This is part of AI For Santa Clarita Businesses: A 2026 Field Guide, the working guide to what AI is actually worth to a business in Santa Clarita.
Connor T. MacIvor · CalDRE #01238257 · Sync Brokerage, Inc. · DRE #02031490