On April 21, someone at Anthropic quietly removed Claude Code from the Pro plan's pricing page. No email, no announcement. People just noticed it was gone. Within a day, Reddit, Hacker News, and X were on fire, and by April 23 it was back. Anthropic's head of growth, Amol Avasare, explained on X that it had been "a small test on approximately 2% of new prosumer signups," and that existing Pro and Max subscribers were never touched.
It was a tiny event. A test that got reversed in two days. But the reason behind it is the interesting part, and it's the same reason surfacing across every AI vendor this spring. Avasare's explanation was blunt: "engagement per subscriber is way up," and "usage has changed a lot and our current plans weren't built for this."
That's the whole story in one sentence. The flat monthly fee was designed for one kind of user, and a different kind of user showed up. For attorneys, CPAs, and consultants building serious AI workflows into client work, that dynamic has direct consequences.
A flat-rate subscription only works if the heavy users and the light users average out. The light users subsidize the heavy ones, the vendor pockets the spread, everyone's happy. That model held for years because most people used AI tools in short bursts. A few questions, a draft, done.
Agents broke the average. When you can point an AI at a task and let it run autonomously for an hour, churning through context and producing tokens the whole time, the "heavy user" isn't 3x the light user anymore. They might be 50x. And there are more of them every month, because running agents is the obvious thing to do once you realize you can.
GitHub said the same thing in different words when it froze new Copilot signups on April 20. Their VP of product noted that "it's now common for a handful of requests to incur costs that exceed the plan price." A handful of requests. Not a heavy month. A handful of requests can cost more than the subscriber pays for the entire month.
When a single user's costs can blow past the price of their plan in an afternoon, flat-rate pricing stops being a pricing strategy and turns into a slow-motion loss the vendor has to stop.
So they stop it. Sometimes with quotas, sometimes by removing a model, sometimes by quietly testing whether they can pull a product off a tier and see who notices. The Anthropic test failed because people noticed fast. The next version will be designed to fail less obviously.
There's a deeper reason this keeps happening, and it has nothing to do with AI specifically. It's the structural problem with subscriptions that people have complained about for years.
One widely shared piece on subscription backlash put the core grievance simply: "When you stop paying, you lose access." The author contrasts it with the old way, where "you bought software, you paid once, you owned it, done." Subscriptions turn software into a lease. That's tolerable when the vendor is genuinely providing ongoing service, like servers or a content library. It's grating when you're renting access to a tool that mostly runs on your own machine and your own work.
For attorneys and CPAs, there's a harder edge to this. It's not just that client work product stored in a vendor's database becomes inaccessible if you stop paying or if the vendor tightens the screws. It's that the client work product was never architecturally yours to begin with. It lived in their system, subject to their data handling practices, their pricing decisions, and their business continuity. That's a different risk profile than a Markdown file sitting in a folder on your hard drive.
And here's the bind: the same agent-driven usage that makes flat rate unprofitable is also what makes vendors most tempted to gate, meter, and reshuffle. The more useful the tool gets, the more pressure there is to take some of that usefulness back.
Even if you ignore the gating, the raw model prices are moving fast enough that planning around any one subscription tier is guesswork. Look at just the last six weeks. Anthropic shipped Claude Opus 4.8 on May 28 at $5 per million input tokens and $25 per million output, holding the line on price from 4.7. Google's Gemini 3.1 Pro sits at $2/$12, well under half of Opus. GPT-5.5 lands at $5/$30.
The spread between those is huge, and it shifts every time someone ships. A subscription locks you to one vendor's pricing decisions. If your tool is wired to Anthropic and Gemini is suddenly less than half the cost for work that doesn't need the top model, you can't act on that. You're paying for the bundle, and the bundle decides which model you get.
What you actually want is the freedom to use the best value model for each task this month, and to do it without the client work you've already produced being entangled with any of it.
I keep coming back to one mental model, and it's held up through every one of these episodes. There are two distinct things you're paying for, and they should never share a fate.
This is the part that genuinely costs money to run, the part that's exposed to all the pricing chaos above. Rent it. Pay the provider directly at cost through an API key, and treat your choice of model as a decision you remake whenever the numbers change. When Gemini's cheaper for a task, use Gemini. When Opus is worth it for the hard matter, pay for Opus on that one. The flexibility is the protection, financially and architecturally. Your API key goes from your machine to the provider. No intermediary sees your client's data.
This is everything you make: the memos, the analyses, the tax position summaries, the engagement deliverables. That should never be rented. If it lives in a vendor's database and your only path to it is their interface, then a pricing crisis on their end becomes an access crisis on yours. You didn't do anything wrong, and your client work is suddenly behind a paywall or a frozen account.
For attorneys, client work product that's only accessible through a vendor's interface is also client work product that's subject to that vendor's subpoena risk, data breach risk, and acquisition risk. The clean version: keep model access cheap and swappable, keep your work in files you own, and let the vendors sort out their economics without taking your client records along for the ride.
This split is the entire premise of Advisor Prep Hero, and the last six weeks have been a decent stress test of whether the premise holds.
The Professional plan is $149/year. You bring your own API key, so model access stays rented directly from the provider, at cost, with no markup from me and nothing for me to gate. When Opus 4.8 shipped, switching to it was a config change. If Gemini's price drops again, that's a config change too. I'm not selling tokens, so the agent-usage problem forcing GitHub and Anthropic to tighten the screws isn't my problem, or yours.
And every document you produce lands as a Markdown file on your machine. No cloud database, no proprietary format, no export step that might break. Your client work is yours regardless of what happens to any vendor's pricing model. That's not a feature I added. It's the part of the design that makes the Anthropic test irrelevant to anyone using it.
Flat-rate AI is going to keep cracking. The agent usage that broke it isn't reversing, and every vendor is going to spend the next year figuring out how to meter, gate, or tier their way back to profitability. Some of those moves will be reasonable. Some will be quiet 2% tests you only hear about when they leak.
You don't have to predict which is which. You just have to stop tying your client work to your tooling. Keep model access cheap and swappable, keep your output in files you own, and let the vendors sort out their economics without taking your practice along for the ride.
Download Advisor Prep Hero free →