If you’re a Claude Code Max subscriber, you probably got an email this week with a credit attached and a vaguely worded apology. If you’re not, you’ve still seen the Reddit threads, the GitHub issues, and the “I burned my entire 5-hour window in 19 minutes” screenshots.

Here’s the operator’s read on what happened, what the credit actually covers, and what it tells you about building on top of someone else’s API.

What broke

Starting around March 23, Max subscribers — the $100 and $200/month tiers — began reporting that their 5-hour session windows were draining in 60 to 90 minutes on workloads that had been fine the week before. One Max 20x user watched a single prompt take them from 21% to 100%. Anthropic acknowledged it on Reddit with the AI-company classic: “We’re aware people are hitting usage limits way faster than expected. We’re actively investigating.”

Two things were happening at once, and Anthropic did a poor job separating them in public.

The first was a deliberate policy change. To manage demand, Anthropic started compressing 5-hour limits during peak hours (roughly 8am–2pm ET) and offering doubled limits off-peak as a sweetener. That’s a reasonable lever to pull. It’s also the kind of thing you announce loudly, not bury in a help-center article.

The second was a bug — or a stack of bugs — in how tokens were being metered. Some users were double-billed through OAuth plugins. Others saw quota math that didn’t reconcile with their actual prompt volume. The community noticed the gap between “we tightened peak limits” and “my single prompt ate my entire window,” and the trust hit was immediate.

The credit, and the OpenClaw twist

Then on April 4, Anthropic dropped the other shoe: Claude subscriptions would no longer cover usage inside third-party harnesses like OpenClaw. If you wanted to keep running Claude through those tools, you’d need a pay-as-you-go bundle or a separate API key.

The credit landed alongside that announcement. Subscribers got a one-time credit equal to their monthly plan cost — $20, $100, or $200 depending on tier — and a link to request a full refund if they wanted out entirely. Anthropic framed it as engineering constraints and sustainable growth. The community read it as “we under-priced agentic coding and now we’re walking it back.”
Both readings are true.

What this actually means if you’re building on Claude

I run automation pipelines on the Anthropic API every day. Signal hits it on a schedule. The consulting work I’m building with my partner depends on it. So this matters to me in a concrete way, and here’s what I took from the last two weeks:

Subscription pricing on agentic tools is structurally unstable right now. Nobody — not Anthropic, not OpenAI, not the harness builders — knows what a fair monthly price for unlimited agent loops actually is. The honest answer is probably “more than $200.” Until the economics settle, expect more quiet tightening, more peak/off-peak gymnastics, and more “we’re investigating” posts.

If your business depends on a fixed-price subscription staying generous, you don’t have a business — you have a subsidy. The people most upset this week were the ones who’d built workflows assuming Max 20x would always cover 24/7 background agents. That was never the deal, even when it worked. If a vendor change of this size would break your unit economics, you need either an API contract or a fallback model.

The API tier is the grown-up tier. It costs more per token, but it’s metered honestly, it’s not subject to surprise harness restrictions, and Anthropic has every incentive to keep it stable because that’s where their enterprise revenue lives. For anything I’m putting in front of a client, I’m running it on the API. The subscription is for my own desk.

Credit is not an apology, it’s a churn-prevention tool. A one-month credit on a $200 plan is $200 of compute that costs Anthropic substantially less than $200 to deliver. It buys them another billing cycle to fix the metering and ship the messaging they should have led with. That’s fine — it’s what I’d do too — but don’t mistake it for the company taking the loss.

The part nobody’s saying out loud

Anthropic has a capacity problem. So does everyone else. The token economy of 2026 is going to look more like cell phone minutes in 2004 than like the all-you-can-eat buffet of 2024, and the providers that communicate that transition clearly are going to keep their power users. The ones that ship policy changes through Reddit replies and help-center edits are going to keep having weeks like this one.
The credit was the right move. The two weeks before the credit were the lesson.
If you’re building anything that depends on a single model provider, this is your reminder to know your fallback, know your true per-task cost on the API, and stop treating any subscription tier as infrastructure.

The AI Ops Playbook goes out every Wednesday at 7am. If a friend forwarded this, you can subscribe at the link below.