If you have been paying $20 or $200 a month for Claude Code, you should probably read this with one eye open.
Over the past year, Anthropic has been quietly tightening the rope on what its subscription plans actually cover. Each move on its own seems reasonable. Stack them together and a pattern shows up that looks a lot like the slow march toward pay-per-token billing for everyone.
Here is what already happened.
In August 2025, Anthropic added weekly rate limits on top of the existing five-hour rolling window. Pro and Max subscribers had been running Claude Code 24/7 in some cases. The cap put a stop to it. The official line was that a tiny fraction of users were burning compute that the rest of the base was paying for. That is probably true. It is also a sign that the flat-rate buffet was already underwater.
Earlier in 2026, things got faster. Anthropic blocked Claude Pro and Max subscriptions from powering OpenClaw and other third-party AI agents. Boris Cherny, head of Claude Code, called it an "outsized strain on our systems." If you wanted to keep using Claude with OpenClaw, you had to top up with usage bundles or switch to a pay-per-token API key. Refunds went out to people who had pre-paid expecting full access.
Then for a little while, brand new Pro subscribers lost access to Claude Code entirely. That one got reversed under public pressure, but the fact that someone at Anthropic clicked the button at all is the part worth thinking about. If your subscription business has a feature your users love, you do not casually try removing it on a Friday afternoon. You only do that if the math is rough enough that it is being seriously discussed in the building.
Enterprise customers got an even rougher deal. Per The Information, Anthropic forced enterprise users off rate-limited subscriptions and onto per-million-token billing in April 2026. Researchers had been finding that some enterprise plans were eating compute worth more than the company was charging. The fix was the obvious one. Charge for what you use, not for the privilege of having an account.
Now stitch it all together. Weekly rate limits. Silent reductions to those rate limits, documented by frustrated users on Reddit who watched their cap drop without warning. Third-party agents pushed off subscriptions. Brief experiments with cutting Claude Code from the Pro tier. Enterprise migrated to metered billing.
What "OAuth subscription getting turned off" might actually look like
The subscription model is not going away in one big announcement. That would be a PR disaster. What is more likely is the same playbook that we just watched play out for OpenClaw, applied to your own day-to-day Claude Code usage:
- The weekly limit gets quietly reduced again. You hit it on a Tuesday. The fix is to "add credits" or "switch to API key billing."
- Specific high-cost features (Opus on long-context tasks, large file edits, autonomous agent loops) start counting against a separate, smaller bucket that resets less often.
- Your one-year OAuth token expires. The renewal flow asks you to upgrade to a "metered Pro" plan that adds a small per-token charge on top of the base subscription. Most people click yes because the alternative is being locked out.
- Enterprise users have already been shoved over. Pro and Max are next once enough of the high-volume users have been priced out and the remaining base looks more sustainable.
None of this is paranoid. Forbes called it "the AI subscription buffet may not last much longer" two weeks ago. Security Boulevard wrote a whole post about how the Claude Code billing backlash is becoming a trust problem. The industry is openly talking about this.
What you can actually do
If Claude Code is part of how you ship work, do not wait for the next "we have updated our plan" email. Hedge now.
Set up an API key alongside your subscription. Claude Code already supports having both a subscription account and an API key configured. You can flip between them depending on what you are doing. Subscription for daily interactive work, API key for batch jobs or anything that would obliterate your weekly cap. The official auth docs walk you through it.
Track your usage. Anthropic does not give you a clean per-prompt or per-token breakdown on the subscription side. You have to estimate it yourself. Knowing roughly what your "weekly equivalent" would cost on the API at retail rates tells you whether the subscription is actually a good deal for you, or whether you are the user the limits were written for.
Have a backup model lined up. The same logic that pushes Anthropic toward metered pricing is going to push every other vendor toward it eventually. Cursor lets you bring your own keys. Continue.dev plus a local model on a 3060-class GPU is not as good as Sonnet 4.6, but it is good enough to keep working when your primary tool stops responding.
Read the renewal email. If your renewal terms change to add per-token charges, you will probably have to actively opt in. Do not just click through.
The honest framing
The flat-rate AI coding subscription was a great deal for power users and a bad deal for the company offering it. Every time inference gets a little cheaper, the heaviest users figure out how to consume even more of it. That gap does not close on its own.
Anthropic is not the bad guy here. Selling a $200/month plan that some users were burning $2,000 of compute on is not sustainable. But the way these changes have been rolling out, with quiet reductions, brief feature removals that get reversed, and refunds for people who already paid, suggests that the company has not figured out how to communicate the shift cleanly. The trust hit is the real cost.
If you are running a business on top of Claude Code, plan for the day when the subscription stops being the default and the API becomes the default. That day is coming. The only open question is how much warning you get before it shows up on your bill.
