At the start of this year the arrangement was understood. American labs held the frontier, Chinese labs shipped cheap alternatives, and if you wanted the best model you paid a US company for API access.

That arrangement is gone, and it did not end with a single dramatic release. It ended with about eight months of relentless shipping while everyone watched the wrong scoreboard.

Chinese labs now hold four of the top five open-weight positions globally. The largest open model ever released is Chinese. A 1.6-trillion-parameter model was trained end to end without a single Nvidia chip, straight through export controls designed to prevent exactly that. Wall Street analysts have started formally recommending Chinese models to clients on price and performance. And when Washington finally reacted, it was to accuse one of those labs of theft and threaten sanctions.

The US still holds the very top. That is the part worth being precise about, because it is true and it is narrower than it was.

Kimi K3 is the largest open model ever built. Moonshot published the full weights in late July, 2.8 trillion parameters, on the Artificial Analysis Intelligence Index scoring 57.1 against Claude Fable 5 at 59.9 and GPT-5.6 Sol at 58.9. That is a gap of a couple of points, not a generation, and it is downloadable.

Alibaba open-sourced at Max scale for the first time. Qwen3.8-Max, 2.4 trillion parameters, published on Hugging Face after a year of Alibaba keeping its strongest models proprietary. The stock rose 7 percent on the announcement.

Zhipu claims to beat Anthropic on security. GLM-5.3 reports 84.5 percent on the CyberGym benchmark against Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6, achieved through post-training alone on an unchanged base model. Zhipu discloses in the same breath that it trails badly on ExploitBench, which is more transparency than most vendors offer.

DeepSeek set the price floor everyone else now gets measured against. V4-Flash beat its own larger flagship on all nine published agent benchmarks, and it natively supports OpenAI's Responses API with Codex adaptation, meaning it drops into OpenAI's own developer tooling as a cheaper engine.

And LongCat was trained without Nvidia. Meituan's 1.6-trillion-parameter model was trained end to end on domestic Chinese silicon, no restricted hardware involved. That is the one that should worry policymakers most, because it means export controls have a ceiling on what they can prevent.

Models topping leaderboards is interesting. Money moving is decisive, and it has started.

Wall Street analysts are now telling clients that Chinese open models deliver most of the frontier's capability at a fraction of the cost, which is a very different thing from a benchmark result. That is a recommendation with fiduciary weight behind it.

The pricing gap explains why. DeepSeek's output pricing has run near $0.44 per million tokens against roughly $30 for GPT-5.6 Sol at points this year. Kimi K3 sits at $3 and $15. Qwen3.8-Max at $2 and $6. Meanwhile Claude Fable 5 is $10 and $50.

Then add the thing an API cannot offer. Weights you can download mean no data leaving your building, no vendor able to deprecate the model underneath you, and no per-token bill at all once the hardware is paid for. For regulated industries and anyone with data residency requirements, that was never available from a US frontier lab at any price, because none of them release frontier weights.

The cadence compounds it. Chinese labs shipped major releases every two to three weeks through the summer. In the same window Google's flagship missed four deadlines, OpenAI paused its largest training run, and Anthropic went nearly a month without releasing anything.

The US response arrived in late July, and it was an accusation rather than a product.

White House science and technology director Michael Kratsios publicly accused Moonshot of distilling Anthropic's Fable to build Kimi K3, alleging a purpose-built internal platform for large-scale extraction against US models. Treasury Secretary Scott Bessent threatened sanctions and Entity List designation, the instrument used against Huawei, with the line that open source is not open season on American IP.

The problem was arithmetic. Fable 5 only returned to public availability on July 1 after its own export-control ban. Kimi K3 launched July 16. Prime Intellect researcher Elie Bakouch said publicly that fifteen days cannot technically account for K3's performance even if distillation occurred. Dean Ball, OpenAI's head of strategic futures, said he does not believe distillation explains the model either, which is an executive at Anthropic's largest rival declining a free shot at a Chinese competitor.

No evidence was published. Anthropic has not claimed it holds any tying K3 to Fable, though it did disclose tracing 3.4 million Claude exchanges to Moonshot back in February.

And then Meta complicated it further. When it open-sourced Muse Glimmer, itself a distilled model, Zuckerberg called for the US to rethink its rules on distillation and training data if it wants to lead in open source AI. The legal line Washington drew against Chinese labs is one American companies now need loosened.

Being accurate here matters more than being dramatic, and the dramatic version is wrong.

The very top of the frontier is still American. Claude Fable 5 and GPT-5.6 Sol both rank above Kimi K3. Astra reportedly solved ten decade-old mathematics problems, with the proofs formalized in Lean so they are machine-checkable, and nothing from a Chinese lab is competing at that level.

Capital is overwhelmingly American. So is enterprise adoption, and so is the infrastructure layer, which is why Nvidia's position matters even to labs training without it.

And every Chinese benchmark claim in this issue is vendor-reported. Moonshot, Alibaba, and Zhipu published their own numbers, in some cases on internal test sets nobody outside can run. The direction is well evidenced. The precise rankings are not.

So the honest read is narrow and still uncomfortable. China did not take the lead. China took the open-model tier outright, closed most of the distance on the closed tier, and did it while shipping faster than every American lab. That is a smaller claim than the headlines and a harder one to argue with.

The stories moving fast beyond today's headlines:

The receipts. Kimi K3 is the largest open model ever at 2.8 trillion parameters, scoring 57.1 on the Artificial Analysis index against Fable 5 at 59.9 and Sol at 58.9. Alibaba open-sourced at Max scale for the first time with Qwen3.8-Max. Zhipu's GLM-5.3 claims a higher CyberGym score than Mythos 5. DeepSeek set the price floor and made its model drop into OpenAI's own Codex tooling. And Meituan's LongCat was trained end to end without a single Nvidia chip.

The money moved. Wall Street analysts now formally recommend Chinese open models on price and performance. DeepSeek output has run near $0.44 per million tokens against roughly $30 for Sol. Downloadable weights also offer something no US frontier lab sells at any price: no data leaving your building, no deprecation risk, no per-token bill.

Washington's response. The White House accused Moonshot of distilling Anthropic's Fable, and Treasury threatened sanctions and Entity List designation. But Fable 5 only returned July 1 and K3 shipped July 16, a fifteen-day window researchers say cannot account for the model, and OpenAI's own head of strategic futures said distillation does not explain it. No evidence was published. Meta then open-sourced a distilled model and asked Washington to loosen the same rules.

What the US still holds. The very top of the frontier, with Fable 5 and Sol both above K3 and Astra reportedly solving ten open mathematics problems with machine-checkable proofs. Plus capital, enterprise adoption, and the infrastructure layer. Every Chinese benchmark in this issue is vendor-reported.

The honest read. China did not take the lead. It took the open tier outright, closed most of the distance on the closed tier, and did it while shipping faster than every American lab. Export controls were meant to prevent this through hardware, and LongCat was trained without that hardware.

Stay building. 🤖