An October 8 SemiAnalysis report burst a comfortable illusion
On October 8, the US research firm SemiAnalysis published "Beijing Will Not Pace the Frontier: China's Speed-First AI Safety Regime." It tallied 857 model releases from nine leading Chinese developers, ByteDance, Alibaba, Tencent, Baidu, DeepSeek, Moonshot, Zhipu, MiniMax, and StepFun, from 2021 through September 15, 2026, and checked each for a published safety evaluation tied to a specific model. The result is stark: only 31 releases, 3.6 percent, ever disclosed one, just 9, 1.1 percent, had it at or before launch, and 813, 94.9 percent, disclosed nothing.
A dialectical cold shower first: the 3.6 percent measures disclosure, not testing
The report is careful with its own wording: it measures public disclosure, not whether companies ran internal safety tests. A model can be tested rigorously in private and simply not publish; conversely, publishing a few flattering results does not mean every risk was checked. So the precise framing is a transparency gap, not a verdict that the models are unsafe. Reuters independently checked the sample scope and method, and the main conclusion holds.
Numbers worth keeping
By company type: of 317 releases from five startups, 20 disclosed, about 6.3 percent, while of 540 from four large firms, only 11 did, about 2 percent. Reasoning models, the think-before-answering kind, were 93 percent undisclosed. The release cadence rose more than thirtyfold, from 3 in Q1 2023 to 101 in Q3 2025, yet quarters with any safety result never exceeded 7. Zhipu is among the few developers that has disclosed safety testing every year since 2022.
The systemic irony: the framework names frontier risks but lands on innovation first
The report flags a contrast: China's own TC260 AI Safety Governance Framework 3.0, dated September 14, names recursive self-improvement, evaluator deception, sandbagging, and shutdown evasion, the exact frontier risks Western labs worry about. Yet the same document puts "encouraging innovation and development" at the highest priority and imposes no capability-based mandatory safety-review duties. In short, China regulates applications and outputs, not frontier risk, moving faster than the EU's compute-threshold approach.
Do not stare only at China: the US side is not spotless either
Dialectically, the same week the US side was not calm. OpenAI had earlier fired three researchers who emphasized safety, and the person who drafted its Preparedness Framework resigned; Anthropic itself blew the internal-eval internet-cutoff story this very week. SemiAnalysis compares disclosure habits and gives no like-for-like US figures, so do not read it as the West is safer. The accurate read is that both sides sit in the same crack where capability outruns governance, just with gaps in different places.
A practical checklist for buying domestic models
For teams putting domestic models into production, three actions: one, ask the vendor for safety notes specific to your scenario, and reject vague claims of being safe; two, keep a human in the loop for critical decisions, especially anything touching money, identity, or outbound messaging; three, retain rollback-capable logs so that when something breaks you can trace which call and which output caused it. None of these depend on vendor disclosure; you can do them yourself, and they hold up best.
What this means for you
For teams using domestic models, the practical takeaway is this: when evaluating whether a Chinese model is fit for production, do not assume it ships with a public safety card, it almost certainly does not. Verify the capability yourself, probe the boundaries yourself, and keep a human in the loop for critical scenarios. Also, do not let the 3.6 percent scare you into thinking the models are unsafe; it measures disclosure habit, not safety level, and large domestic labs may test privately without saying so. What you should track in every case, home or abroad, is whether the model you use has someone to reach and a source to trace when something goes wrong.