Avoid Blind Upgrades: Claude Code 2.1.81 May Multiply Token Usage

If you recently upgraded to Claude Code 2.1.81 you may notice three symptoms: token consumption surges dramatically, the conversation context frequently "forgets", and the same task costs several times more.

Under normal operation Claude Code reuses cached context to keep token usage low. In version 2.1.81 a critical exception occurs: the cache creation fails, the context cannot be hit, and every request is recomputed from scratch, causing token usage to increase multiplicatively.

Cache creation failure → Context miss → Full‑round computation each request → Token usage multiplies

We illustrate the failure with a flow‑chart:

The problem is dangerous because it does not raise an error and does not affect functionality, yet it silently burns tokens.

The bug only impacts non‑official integration paths. If you use Anthropic’s official API you are unaffected. Affected scenarios include local inference services (self‑hosted proxies), third‑party API gateways, and any unofficial access method. Version 2.1.81 introduced extra headers and a new cache strategy that these paths do not handle, leading to cache invalidation.

Temporary fix: downgrade to the stable 2.1.77 version, which restores the reliable cache mechanism. npm i -g @anthropic-ai/[email protected] To prevent automatic upgrades, set the environment variable CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=true (add it to your shell profile or .env file on servers).

export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=true

Additionally, disable the extra attribution header by editing config.json:

{
  "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
}

These changes do not erase existing configuration or memory; the downgrade merely replaces the runtime binary while preserving stored data.

The incident illustrates a broader engineering lesson: cache mechanisms are hidden infrastructure that, when broken, are hard to detect but can dramatically increase cost and degrade experience. Treat version upgrades with risk awareness rather than assuming newer is always better.