 |
VOOZH | about |
|
VOOZH | about |
If you use Claude Code and have noticed in recent weeks that your usage limits are running out much faster than usual, it's not your imagination: it's a confirmed bug. A security researcher reverse-engineered the Claude Code binary and discovered two independent bugs in the prompt cache that silently multiply token consumption by up to 20 times.
Anthropic has acknowledged the issue. Lydia Hallie, product lead of Claude Code, wrote on X: "We're aware people are hitting usage limits in Claude Code way faster than expected. Actively investigating."
Same-day twin announcements: Anthropic raises weekly caps by 50% through July 13 to stop individual dev migration. OpenAI counters with two free months of Codex for enterprises switching within 30 days. Frontal battle on the AI coding segment.
→ Read the full battle analysisLive now for Pro, Max, Team and seat-based Enterprise. Defensive move: many users are migrating to OpenAI Codex for lower token consumption and comparable performance.
→ Read the full analysisOver 500 users have reported the issue on GitHub. The impact is devastating:
Financial impact: If you pay per usage via API, these bugs can inflate your bill by 10-20x with no warning whatsoever. Every request that should read tokens from the cache is instead fully reprocessed from scratch.
Researcher skibidi-toaleta-2137 on Reddit explained that they spent days reverse-engineering the standalone Claude Code binary (228 MB, ELF format) using Ghidra, a MITM proxy, and radare2. They found two independent bugs that break the prompt cache.
The standalone Claude Code binary uses a custom fork of Bun (the JavaScript runtime acquired by Anthropic in 2024). This runtime performs a string substitution on every API request, looking for a special billing identifier.
The problem: if the chat history mentions any billing-related content, the substitution hits the wrong location and breaks the cache prefix. The result is that every subsequent request is processed from scratch, without benefiting from the cache.
Switch from the standalone version to the npm version. Instead of using the downloaded binary, run:npx @anthropic-ai/claude-code
This bypasses the custom Bun runtime and uses standard Node.js, where the bug does not exist.
Using the --resume or --continue flags to resume a previous session causes a cache prefix mismatch. Instead of reading the conversation from the cache, the system rewrites the entire history on every turn.
This bug was introduced in version 2.1.69 and has not yet been officially resolved.
No clean fix for Bug 2. The only current workaround is to avoid using --resume and --continue. Start new sessions instead of resuming old ones.
Lydia Hallie (@lydiahallie), product lead of Claude Code at Anthropic, confirmed on March 31, 2026 on X (formerly Twitter):
"We're aware people are hitting usage limits in Claude Code way faster than expected. Actively investigating, will share more when we have an update!"
Source: Post on X - March 31, 2026, 432,500 views
The post received 766 replies, 529 reposts, and 6,200 likes, a testament to the widespread frustration in the community. Anthropic stated this is the "team's highest priority".
The bug was filed on GitHub as Issue #40524 and was classified as a regression bug and assigned for resolution.
The situation is complicated by the fact that there are two overlapping issues:
This means that even after the cache bugs are fixed, limits will still be more restrictive than they were before March 26.
Multiple users have confirmed that downgrading to 2.1.34 makes "a very noticeable difference" in token consumption.
Do not resume previous sessions. Always start a new session and, if necessary, pass context manually.
Cloud MCP connectors (such as Ahrefs, Supabase, Similarweb) inject their full tool definition schemas into every API call, even when you're not actively using them. Remove any connectors you don't actively need.
Use the /cost command during a session. If you see that input tokens grow massively with each message (instead of remaining stable thanks to the cache), the bug is affecting you. Another sign is reaching your usage limit in a much shorter time than usual.
Both. Pro/Max users see their quota drain faster. API users are paying bills 10-20x higher than expected. The mechanism is the same: the prompt cache is not working, so every message is reprocessed from scratch.
No announcement has been made yet. Anthropic has acknowledged the problem and is investigating it as their highest priority, but has not mentioned refunds. If you use the pay-per-use API and have noticed abnormal bills, contact Anthropic support with your session logs.
Yes, for Bug 1. Switching to npx @anthropic-ai/claude-code bypasses the custom Bun runtime that causes the cache problem. Multiple users on GitHub have confirmed a significant improvement. It does not fix Bug 2 (--resume).
There is no date. Issue #40524 on GitHub has been assigned and classified as a regression, which means it is being worked on. Lydia Hallie has promised updates. In the meantime, use the workarounds described above.
Through reverse engineering. A researcher (skibidi-toaleta-2137 on Reddit) analyzed the 228 MB ELF binary of Claude Code using Ghidra, a MITM proxy, and radare2. They intercepted the API traffic and noticed that requests never used the cache, rewriting the entire context on every turn.
This bug is a serious problem for anyone who uses Claude Code on a daily basis. The good news is that Anthropic has acknowledged it and is treating it as their highest priority. The bad news is that, at the moment, there is no complete official fix.
Immediate actions: Use npx @anthropic-ai/claude-code, avoid --resume, monitor with /cost, and remove unused MCP connectors. If costs are abnormal, contact Anthropic support.
Want to stay updated on this bug or need help with the workarounds? Use the contact form below to get in touch.
Subscribe to the newsletter to receive new articles directly in your inbox.
Subscribe to the newsletter to receive new articles directly in your inbox.
3.4k readers worldwide, every Saturday