Skip to content
CC Switch WikiUnofficial Community Wiki

x-anthropic-billing-header breaks caching - CC Switch

Quick fix

Upgrade CC Switch to v3.15.0+, which fixes the dynamic billing header breaking prefix caching in openai_responses mode.

Symptom

Section titled “Symptom”
Error output
cached input tokens were often stuck around 6.5k-9.2k even for ~95k-120k token requests

Cause

Section titled “Cause”

This cluster actually mixes several quite different CC Switch request conversion and proxy defects. This page addresses the fixed core problem (#2350): in `openai_responses` mode, CC Switch incorrectly injected an `x-anthropic-billing-header` containing a dynamic `cch` value at the start of the `instructions` field. Because OpenAI prompt caching matches on a prefix, that header changing on every request stopped a large, otherwise stable system prompt from ever hitting the cache.

The cluster also holds unfixed variants: Azure OpenAI `max_tokens` not being converted to `max_completion_tokens`, producing a 400 (#3185); `supports_reasoning_effort()` using `strip_prefix("gpt-")` so a model with a custom prefix never triggers reasoning mapping (#5343); Claude Desktop routing not reusing the Session ID, lowering the cache rate (#5388); and `<thinking>` tags leaking into the body on the Codex chat-completions route (#6116).

Fix

Section titled “Fix”

  1. Upgrade CC Switch to v3.15.0 or newer. That version fixed the dynamic `x-anthropic-billing-header` being injected into the `instructions` field in `openai_responses` mode, restoring the OpenAI prefix cache hit rate.

Affected Versions

Section titled “Affected Versions”
ToolClaude Code
Version3.14.1 (已在 3.15.0 修复)
PlatformsWindowsmacOS

Source Issues

Section titled “Source Issues”

This page is distilled from 6 real issues

FAQ

Section titled “FAQ”
Does a quota refresh use the proxy from the config file environment variables, or the outbound proxy in the app?
A quota refresh uses the CC Switch HTTP client: the global outbound proxy if one is set, otherwise the system proxy. A provider's `env` proxy settings are not used.
Why does a custom model name with a prefix (b-gpt-5.6-sol, say) not trigger reasoning effort mapping?
The current `supports_reasoning_effort()` matches with `strip_prefix("gpt-")`, requiring the model name to start with `gpt-`. A model with a custom prefix is not recognised, and this is not fixed yet.
How do I resolve Azure OpenAI gpt-5.4 returning a 400?
CC Switch still sends the old `max_tokens` parameter while Azure GPT-5.x requires `max_completion_tokens`. That parameter conversion is not fixed in CC Switch yet — watch for a later version.
Section titled “Related problems”
ANTHROPIC_BASE_URL rewritten to 127.0.0.1With local routing on, ANTHROPIC_BASE_URL pointing at http://127.0.0.1:15721 is expected. Move shared fields into 通用配置 and restart Claude Code.10 issues merged"Expecting property name enclosed in doubleUpgrade CC Switch to v3.16.4 or newer to fix the JSON parse error caused by a large request body not being decompressed.2 issues mergedYour access token could not be refreshedUpgrade cc-switch to v3.17.0+, which fixes the PROXY_MANAGED placeholder polluting OAuth credentials or causing a 401 during takeover.5 issues merged

This is an unofficial community wiki with no affiliation to the cc-switch authors or the project itself. Its content is compiled from the project's public GitHub issues. This site distributes no software.