Opus 4.8, Qwen, DeepSeek, and a Claude Code Failure: What I Could Actually Reproduce

#ai #claude #llm #testing

There is a claim going around that Claude's latest Opus 4.8 may have been distilled from Qwen or DeepSeek.

That kind of claim spreads quickly, especially when it can be turned into a screenshot or a short clip. I wanted to test the small version of the claim first: if I ask Opus 4.8 what model it is, does it identify itself as Qwen or DeepSeek?

In my May 30 test, I could not reproduce that behavior.

But the more useful part of the test happened before the model test even started. Claude Code broke after an upgrade with spawn EBUSY, and Codex helped diagnose and fix the local Claude Code state.

The first failure was not the model

I originally planned to open Claude Code, switch to the new Opus 4.8 path, and ask a direct identity question.

Instead, Claude Code failed after the upgrade with:

spawn EBUSY

This is the kind of problem that is easy to misread. When an AI coding tool fails to start, it is tempting to blame the account, the network, the subscription, or the remote model service.

Codex pointed in a more local direction: the Claude Code component state looked broken.

The useful clues were:

an old session file parsing problem
a Claude Code executable that appeared to be half-downloaded, locked, or otherwise incomplete

After cleaning up the local component state, Claude Code ran again.

This is a very normal kind of AI tooling failure. The demo version of AI coding looks smooth. The real version often includes local caches, CLI updates, broken sessions, locked binaries, and confusing error messages.

If the toolchain is broken, the model has not really been tested yet.

Then I tested the identity claim

After Claude Code was working again, I asked Opus 4.8 a direct question:

What large model are you?

In this run, it identified itself as Claude Opus 4.8, developed by Anthropic, and running in the Claude Code environment.

It did not identify itself as Qwen.

It did not identify itself as DeepSeek.

The careful conclusion is:

In this test material, I did not reproduce Opus 4.8 identifying itself as Qwen or DeepSeek.

That is intentionally narrow.

It does not prove anything broad about training lineage, distillation, data contamination, or evaluation artifacts. A single self-identity answer is not a rigorous method for determining model origin.

But it does mean I would not treat the stronger viral claim as settled without more reproducible evidence.

The practical lesson: keep more than one agent

The most useful part of this test was not the model identity answer. It was the workflow lesson.

Claude Code broke. Codex helped fix Claude Code.

That suggests a practical setup for anyone using AI coding tools seriously: keep more than one agent installed.

For example:

If Claude Code fails, ask Codex to inspect logs and local state.
If Codex hits a confusing error, ask Claude to analyze the message.
If one toolchain is stuck, use the other agent to preserve diagnostic momentum.

This is not about declaring one tool better than another. It is about avoiding a single point of failure in your AI workflow.

Do not outsource verification to the timeline

The second lesson is about model rumors.

Claims like "this model is distilled from that model" or "this model is just a wrapper" are easy to share. They may be worth investigating, but they should not be accepted from a screenshot alone.

A better habit is to record: