Okay, About That Claude Thing I Said I'd Follow Up On

I said I'd follow up. Here's the follow up.

If you listened to Episode 005 (listen here), you heard David absolutely lose his mind about Claude burning through credits mid-workflow and then asking for more money to finish the job.

I also said I'd post a full breakdown once I had actual answers. So here we are.

Here's what's actually going on.

The short version: it's real, it's widespread, and Anthropic knows.

This isn't a you problem. It isn't a "you're using it wrong" problem. The credit drain people have been experiencing -- sessions maxing out waaaaaay faster than expected, usage meters jumping to 100% on what should be routine tasks, daily limits gone before the actual work gets done -- is happening at scale, and Anthropic has publicly acknowledged it. Finally.

Their statement: people are hitting usage limits in Claude way faster than expected, and they're actively investigating. They called it their top priority.

That's a start. It's not nothing.

But here's the thing -- it's actually a few different problems stacked on top of each other.

Problem one: Anthropic tightened session limits during peak hours. Something I took note of: this occurred during a time that they had also advertised that session limits were doubled for a limited time, yet what was happening was the exact opposite. They've said the weekly total is unchanged, but the per-session ceiling got lower. If you do most of your AI work during normal business hours, you're working in peak hours. That matters.

Problem two: there's a suspected bug in the prompt caching system that may be inflating token counts -- meaning Claude is logging more usage than it's actually doing. Reports suggest this could be running 10 to 20 times over actual consumption in some cases. That's not a rounding error. That's a billing problem.

Problem three: the limits themselves aren't clearly documented. The Pro plan promises "at least 5x the usage of the free plan." The Max plan promises "5x or 20x the Pro plan." What does that actually mean in practice? Nobody outside Anthropic knows. And that lack of transparency is its own issue -- because when you don't know what you're supposed to get, you can't tell when you're being shorted.

So what do you actually do right now?

Work outside peak hours if you can. Early morning and evenings tend to have more headroom than standard business hours -- that's when session limits are less compressed.

Break big tasks into smaller defined steps instead of one long "build me this whole thing" prompt. Multi-step workflows where Claude is making multiple tool calls back to back are some of the heaviest credit consumers. Chunking it down gives you more control and more visibility into where your usage is going.

If you're running Claude in any kind of automated workflow -- and some of you are -- know that rate limit errors can look like generic failures and silently trigger retries. One loop can drain your entire daily budget in minutes. This is a known issue. Build in explicit error handling.

And if your credits are disappearing faster than they should be on a task that genuinely seems simple -- that may be the caching bug, not your prompting. It's worth knowing that's a possibility before you assume you're doing something wrong.

Is this a dealbreaker?

For me? Not yet. But I'm watching.

The lab stays honest. Claude earned a real review when it earned it. It's getting a real follow-up because the situation changed -- not because the opinion changed. There's a difference between a product hitting a rough patch and a product being fundamentally dishonest about what it delivers.

Right now this reads as growing pains at scale. A lot of demand, not enough infrastructure to absorb it cleanly, limits that got tightened without enough communication. That's forgivable -- IF it gets fixed and if Anthropic actually tells people what's happening in real time instead of waiting for Reddit to blow up first.

The silence is the part I'm still annoyed about.

Astro knocked the Claude browser tab closed twice while I was writing this. I'm choosing to believe that's editorial feedback and not a coincidence.

Experiment continues. I'll update this post if anything significant changes.

-- Annie 🧪

Have you been hitting credit limits faster than expected? Drop it in the comments, or tell me about it a the  Contact Us page. Real data from real users is how we actually track whether this is getting better or worse.

Related: ChatGPT vs. Claude -- the full Lab Report | AI Alchemy with Annie podcast