In late February 2026, a Claude Code release shipped with a bug in prompt caching. For about a day, usage limits drained faster than normal, and users had no way to tell. Many blamed their own habits; some considered paying for a bigger plan. Anthropic fixed the bug within a day and reset everyone's limits. But the episode exposed a risk that goes well beyond one incident.
What happened
Prompt caching lets a model reuse the parts of a conversation it has already read, such as the system prompt, the tool definitions and the earlier turns, instead of paying full price to read them again on every request. In a long coding session most of each request is that repeated context, so caching is a large part of what keeps the cost down.
When caching breaks, that repeated context is billed as fresh input on every request. Nothing looks different on screen and the answers are the same. The only symptom is that usage climbs faster.
On February 26, 2026, a member of Anthropic's Claude Code team posted on X:
"We've reset rate limits for all Claude Code users. Yesterday we rolled out a bug with prompt caching that caused usage limits to be consumed faster than normal. This is hotfixed in 2.1.62."
- Around February 25A Claude Code release with a prompt-caching bug ships.
- About a dayUsage drains faster than normal. Nothing on screen shows why.
- February 26A hotfix (2.1.62) ships and everyone's limits are reset.
The self-blame problem
In the hours before the bug was found, people who hit their limits had no way to know anything was wrong. The Claude interface shows how much of a limit is used. It doesn't show the tokens each interaction consumed, the split between input and output, or whether caching is working.
So people assumed they had overdone it. Some held back. Some considered upgrading. Some stopped working until their limit reset. A platform bug became each user's problem.
The reset helped, unevenly
Resetting everyone's limits was a fast, generous response. But a blanket reset can't undo what an incident like this costs. People who held back lost the work they didn't do. Anyone who didn't update straight away kept draining faster until they did. And on a weekly limit, a reset lands differently depending on where each person was in their cycle.
Software that meters itself will have bugs. Caching, retries, duplicate requests and context handling are all places where a small regression shows up only as a bigger number.
One source of truth
The incident exposed a structural risk: relying on the AI vendor for both the service and the measurement of that service.
If your organization uses Claude across teams, ask whether you can see:
- Tokens per interaction: how much each request consumed.
- Input, output and cached tokens: where the money actually went.
- Whether caching is working: the difference between normal and broken.
- Per-person patterns: who uses which model, and how heavily.
- A baseline: whether today looks like a normal day.
The core insight: when the same vendor provides the service and the only view into it, there's no independent check. A bug that inflates your costs looks exactly like legitimate heavy use. The only party that can tell the difference is the one sending the bill.
Would you know?
If you lead FinOps, IT or engineering, the February incident raises one question: would you have known if this happened to your team?
Not whether the vendor would eventually fix it. They did, within a day. The question is whether you'd have had any signal at all before the vendor said so.
On a subscription plan, a bug like this makes people hit their limit early and stop. On usage-based billing there's no limit to hit unless you set one: a caching failure that doubles consumption across a team for even a day turns straight into money.
Keep your own record
You wouldn't run cloud infrastructure on the provider's word alone. You watch your cloud spend in your own monitoring, and you reconcile what your payment processor reports. AI tools are no different: the spend is real and growing, and the vendor's dashboard can't tell you when the vendor's software is wrong.
An independent record of AI usage, taken between your teams and the AI providers, lets you:
- See tokens and cost per request, by person and by model, not just a percentage bar.
- Check that caching works: cached tokens show up in each request's record, or they don't.
- Reconcile the bill against your own numbers, side by side with what the vendor reports.
Where ZuzAI fits
ZuzAI keeps your own record of AI usage: tokens and cost per request, by person, tool and model, next to the cost each vendor reports through its admin API. It runs inside your own tenant.
Published by ZuzAI on June 30, 2026.
