Solving the Claude Code API Error 400 Deferred Problem

Reproducing the Problem

Recently I connected Claude Code to z-ai/glm-5.3-flash on OpenRouter and hit two pitfalls along the way. My money is already topped up on OpenRouter and I didn't want to move it to Z.ai. So I want to write down my complete setup and the configuration that finally worked. Maybe it can help someone.

To make comparison easier, let me first post my earlier cc-switch configuration here:

The places to focus on are already circled in red, mainly the request URL and the upstream format configuration. If you configured it this way too — and of course this is also the fastest way, nearly just filling in a key and you can use it — then as you keep using Claude Code you will very likely run into the problem I ran into:

PLAINTEXT

复制

 API Error: 400 Deferred custom tools are only supported on Anthropic models
 and on Anthropic-compatible provider endpoints that implement deferral.
 Other endpoints cannot call tools omitted from tools[].
 Received z-ai/glm-5.3-flash-20260826

The cause is that Claude Code marks some tools as deferred (Tool Search / on-demand loading): the request can leave those tools out of tools[] at first and pull them in later with tool_reference. This is an Anthropic-specific capability.

OpenRouter sees that the target model is GLM — not Claude, and not an "Anthropic-compatible endpoint that implements deferral" — and rejects it outright. The same kind of error appeared earlier with stealth/ox-alpha. Ox Alpha was later officially identified as GLM-5.3-Flash, so this isn't a one-off but a protocol mismatch.

+

The Solution

I won't go into the obvious solutions here, such as switching directly to the endpoint with the Anthropic protocol that GLM officially supports and connecting through GLM's official API. Of course, I haven't verified that approach myself, because my money is all topped up on OpenRouter and I couldn't be bothered to use the official API.

The method I used is cc-switch's local routing: open cc-switch's settings screen, choose Routing, and the reference configuration is as follows:

This is step one; we also need to change the model configuration in cc-switch:

• Request URL: https://openrouter.ai/api/v1

• Upstream format: OpenAI Chat Completions (routing must be enabled)

That's roughly the flow. After configuring it, reopen Claude and the problem should be solved — at least it solved mine. So I wrote this post to share it....

+