Chinese AI Closes Gap with Anthropic After Claude Code Source Leak
Claire Weston
A packaging error exposed 513,000 lines of Claude Code's full client-side source; within four months three Chinese labs released models approaching Anthropic's flagship — shifting the competitive focus from model scale to the engineering scaffold built around the model.
What exactly leaked?
Anthropic accidentally shipped a 59.8 MB source-map file in npm package version 2.1.88, reconstructing roughly 513,000 lines of unobfuscated TypeScript across about 1,900 files.
No model weights, training data, or user data were involved — Anthropic called it a packaging error. What leaked was the full client-side "harness" — the engineering layer that wraps the model.
This means → outsiders saw, for the first time, exactly how a top lab turns a large model into a usable product: memory systems, permission gating, multi-agent coordination, and system-prompt assembly.
Why is the "scaffold" worth more than the model itself?
The developer community converged on a single conclusion: the real value sits not in the model but in the engineering layer around it — multi-tier memory, permission-gated tools, multi-agent orchestration patterns.
In plain terms = the model is the engine; the scaffold is the chassis and electronics. A powerful engine without a chassis cannot drive.
"The scaffold is the new moat" became the prevailing judgment. Industry discourse shifted from "writing prompts" to "writing operating systems."
What did the three Chinese labs ship?
Zhipu AI (Z.ai) released GLM-5.2 on June 16 — a mixture-of-experts model with roughly 750 billion parameters, a 1-million-token context window, and an MIT license. It leads Fable 5 by about 10 Elo points on the Design Arena coding leaderboard.
Moonshot AI released Kimi K3 in mid-July at 2.8 trillion parameters, positioned as the largest open-weight model to date. Across 35 benchmarks, Fable 5 won 22, Kimi K3 won 12, and 1 was a tie — performance described as "within Fable 5's range."
Alibaba previewed Qwen 3.8-Max on July 19 at 2.4 trillion parameters, trailing Fable 5 by roughly 12.6 points in a benchmark aggregate; no full-release date has been announced.
Did they use the leaked code?
No evidence supports the claim that any of the three models directly benefited from the leaked code. All three were in development before the leak, each using an independent architecture.
Kimi K3, for instance, runs on Moonshot's proprietary Kimi Delta Attention hybrid linear-attention mechanism — nothing suggests derivation from Anthropic's system.
This reflects a documented convergence: all three labs independently moved toward the same engineering layer the leak revealed — Zhipu shipped a dedicated scaffold called ZCode alongside GLM-5.2, and Moonshot runs Kimi K3 on its own KimiCode scaffold.
What happened on Anthropic's side?
Anthropic released Fable 5 on June 9, priced at $10 per million input tokens and $50 per million output tokens.
The model was briefly taken offline on June 12 under U.S. Commerce Department export restrictions; the restrictions were lifted June 30 and Fable 5 redeployed on July 1.
Anthropic has previously accused several Chinese firms — including DeepSeek, MiniMax, and Moonshot — of distilling Claude outputs via fake accounts. The allegation remains unverified; the companies have not acknowledged it.
What should we watch next?
The competitive focus has shifted from "whose model is bigger" to "who can turn a model into a reliable, production-grade agent" — precisely the capability the leak inadvertently exposed.
This means → for investors, the key variable is no longer parameter count or benchmark scores alone, but whether each lab can sustain its scaffold-level engineering advantage.
In plain terms = the engines are converging; the next race is about who builds the better car.
Content is for reference only, not financial advice.