Claude Code fixes cache hit bug: user fees once surged several times due to cache invalidation

One sentence conclusion

Claude Code fixed a caching bug that caused user fees to soar several times, which is a major benefit for teams that rely on AI programming tools for automated development. But this incident exposed the problem of cost transparency of AI productivity tools: users need a clearer expense early warning mechanism instead of discovering billing abnormalities after the fact.

What happened

In April 2026, Anthropic's AI programming tool Claude Code released version v2.1.90, which mainly fixed a serious cache hit bug. The bug exists since v2.1.69 and when a user resumes a previous programming session using the --resume parameter, the first request triggers a full cache hit instead of using cache optimizations as expected.

This means that when the developer resumes work, the first request will be treated as a "new calculation" by the system and consume the full token quota. For heavy users, this bug could cause monthly charges to spike several times, with the user not realizing the problem until they receive their bill.

The technical community discovered this issue in a source code leak, and Anthropic quickly released a repair version. In addition to this major bug, v2.1.90 also fixed several other issues that affected the user experience, including delayed interface response and loss of some functions.

Direct impact on practitioners

Cost level

For developers who use Claude Code for 1-2 hours a day, the monthly cost was originally between 343-687 yuan. However, affected by this bug, the actual cost may soar to 1373-4119 yuan/month, and heavy users may even exceed 13731 yuan/month.

After the repair, costs return to normal levels, but users need to re-evaluate their usage patterns: after enabling advanced features such as Agent Teams and scheduled tasks, costs will still increase significantly.

Efficiency/capability level

The cache bug fix means developers no longer need to worry about the "first request penalty" when getting back to work. This is especially important for teams that need to switch projects frequently or handle multiple parallel tasks.

But what is more noteworthy is that this incident exposed a common problem with AI productivity tools: it is difficult for users to predict and control the actual cost of use. When tools automate complex tasks, expenses can accumulate inadvertently.

Core data comparison

Our judgment

**The warning of this bug fix is underestimated. **

On the surface, this is just a bug fixed by the technical team. But looking deeper, it reveals a key weakness in the AI ​​productivity tool ecosystem: insufficient cost transparency.

Reason 1: The billing model of AI tools is too complicated. When tools automatically perform tasks, it is difficult for users to monitor token consumption in real time. The caching bug of Claude Code lasted for several versions before being discovered, which shows that the existing monitoring mechanism is not effective enough.

Reason 2: Tool providers have insufficient early warning of cost risks. If users were alerted immediately when charges were abnormal, rather than waiting until the end of the month for a bill, the severity of the problem would be greatly reduced.

Those who benefit the most are small and medium-sized development teams and independent developers, who are more cost-sensitive but lack dedicated financial monitoring tools. The least affected are large enterprise users, who have specialized budgeting and monitoring systems.

This incident should prompt all AI tool providers to rethink how to provide powerful functionality while ensuring that users do not face financial risks due to technical details.

3 things you can do now

1. Check the current version: Confirm that your Claude Code has been updated to v2.1.90 or higher, run claude-code --version in the terminal to check.

2. Set usage reminder: Enable token usage reminder in Claude Code configuration, or use third-party monitoring tools to track daily consumption.

3. Evaluate Alternatives: If your project is extremely cost-sensitive, consider testing the cost performance of other AI programming tools and establishing alternatives.

Related tools (jump within the site)

This article involves tools: Claude Code