The AI Brief

Speaking to reporters at his Doonbeg golf resort in Ireland on Sunday, President Trump dismissed the pacing call from Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and xAI CEO Elon Musk as the work of "negative forces" making claims about things that won't happen. "We're leading China in AI, we're the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI wins," Trump said. He added that the US could "put guardrails" but offered no specifics on regulatory measures.

The president's comments came three days after Amodei published "We Must Pace the Frontier," a roughly 3,800-word essay on September 12 proposing embedded third-party evaluators and cross-lab safety checkpoints. Altman endorsed the framework on X within hours; Musk followed with three words: "Dario is right." Google DeepMind chief Demis Hassabis also concurred. The speed of that consensus among ordinarily competing executives was itself unusual; the White House's speed of rejection was equally fast.

Former White House AI adviser David Sacks, now co-chair of the President's Council of Advisors on Science and Technology, offered a narrower rebuff: labs can slow down on their own, but seeking an antitrust waiver or a new government approval regime as the price of doing so amounts to "blackmail" of the political system. House Speaker Mike Johnson called for meetings rather than legislation. Democratic leaders pushed back on the White House's posture, with Arizona Senator Ruben Gallego warning that dangerous AI systems will not respect national borders if safety guardrails fail.

Japan's Nikkei Shimbun, citing OpenRouter data, reported that Chinese AI accounted for 55% of global token usage on the platform in June 2026, compared with 43% for US-origin models. US labs held 76% of the same traffic a year earlier. As of September 7, Tencent's Hy4 Preview had taken OpenRouter's top position as the single most-used model, per Seoul Economic Daily. DeepSeek had previously held that position with 17.6% of routed tokens.

Pricing is the primary mechanism. DeepSeek V4-Flash costs $0.14 per million input tokens; OpenAI's GPT-5.6 Sol is $4.00 per million input. Tencent Hy4 Preview is priced at roughly $0.83 per million input tokens through Tencent Cloud TokenHub and OpenRouter. OpenRouter COO Chris Clark noted that Chinese open-weight models are "disproportionately heavy in agentic flows run by US firms," meaning the shift is not confined to experimental or hobbyist usage.

The shift also accelerated the supply concentration risk that some enterprise buyers have begun to price in: Anthropic's share on OpenRouter fell from 29.1% to 13.3% over twelve months, per figures tracked by aggregator Dentro, with six Chinese models now ranking above Claude. Operators self-hosting Chinese open weights under permissive Apache 2.0 licenses can bypass data-routing exposure, but the 770B parameter footprint of models like Hy4 requires roughly 924 GB of storage even in FP8 quantization, putting local deployment out of reach for most teams.

Wired reported that OpenAI approached members of Congress in recent weeks to ask whether orchestrating an industry-wide pause in frontier AI development would violate federal antitrust law. The core anxiety, confirmed by people close to the company, is that a formal pact among OpenAI, Anthropic, and Google to coordinate model releases or cap compute scaling could constitute jointly restricting output, a potential Sherman Act violation. OpenAI CEO Sam Altman told employees at an all-hands that the company is willing to moderate its release cadence if peers agree to synchronized restraint, but acknowledged deep skepticism about compliance.

The only legislative vehicle to address this is S.5105, the Collaboration on Adversarial Threats and Security Risks Act, introduced by Senators Adam Schiff and Jim Banks on July 23, 2026, with a companion House bill. The legislation would exempt inter-company agreements from antitrust rules when they aim exclusively to reduce risks tied to advanced AI models, and would explicitly permit coordinated delays or limits on deployment, training, or evaluation of AI systems. The bill has not advanced to a floor vote.

Former White House AI adviser David Sacks made the structural problem explicit: labs can already slow down unilaterally without any legislative action, but conditioning a voluntary pause on antitrust immunity or a new regulatory approval regime is what he called "blackmail" of the political system. The legal uncertainty, even absent a lawsuit, acts as a deterrent to formal coordination, pushing any real pacing framework toward either unilateral action by individual labs or waiting for Congress to act on S.5105.

Joe Benton, who led a safety research team at Anthropic focused on scalable oversight, and Josh Engels, an AI safety researcher at Google DeepMind, gave their first public interviews since leaving their positions to NBC News, published September 9-10, 2026. Both said they are joining METR, the nonprofit AI evaluation organization that independently investigated the Hugging Face breach in July 2026. Engels said he had turned down offers from both Anthropic and OpenAI before choosing METR.

Their departures follow Jacob Coxon's viral resignation from Anthropic on September 8, in which Coxon, who previously contributed to GPT-4o at OpenAI, accused both companies of "gambling with our lives" by moving toward self-improving superintelligence without adequate safeguards. Benton's concern is more specific than Coxon's: he said the absence of any mandatory reporting requirement means the public cannot track how often frontier systems have already breached human instructions, and that the gap will widen as models become more capable. Engels told NBC: "There are no adults in the room."

Anthropic said the company "has always been transparent that AI will bring both enormous benefits and unprecedented risks" and that it builds models with "some of the strongest safeguards in the industry." OpenAI said newer public models, including Astra, more reliably follow human instructions and that the company has strengthened its safeguards. METR gaining two senior researchers adds to the independent evaluation infrastructure: Coefficient Giving's $200 million push for external safety organizations is separately underway, and METR's eight-week Anthropic audit, signed in September, remains in progress.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Tencent released and open-sourced Hy4 Preview on August 28, 2026. The model uses a Mixture-of-Experts architecture with 770 billion total parameters and 49 billion active per token, a one-million-token context window, and an Apache 2.0 license covering the BF16 and FP8 checkpoints distributed through Hugging Face, ModelScope, GitCode, and GitHub. Tencent also published vLLM and SGLang deployment guides, fine-tuning instructions, and an AngelSlim compression toolkit. As of September 7, Hy4 Preview ranked as the most-used model on OpenRouter, displacing DeepSeek from the top slot.

On Tencent's own internal blind evaluation, 163 experts scoring 203 engineering tasks rated Hy4 Preview at 2.99 out of 4, versus 2.92 for GLM-5.3 and 2.94 for Kimi K3. On public benchmarks it generally trails Claude Opus 5 and GPT-5.6 Sol, trading wins with other Chinese flagships. The pricing advantage is the operative differentiator: at $0.83 per million input tokens through Tencent Cloud TokenHub, Hy4 Preview undercuts OpenAI GPT-5.6 Sol ($4.00/MTok) by roughly 5x and closed Anthropic Opus 5 ($5.00/MTok) by roughly 6x, while remaining permissively licensed.

The data-sovereignty wrinkle is material. API calls routed through Tencent Cloud are governed by China's Cybersecurity Law, amended January 2026 to explicitly include cloud services. Teams that download and self-host the open weights under Apache 2.0 sidestep that routing risk entirely, but the weights require approximately 924 GB of storage in FP8 quantization, making local deployment a data-center project. OpenRouter's measured throughput for the model at launch was roughly 35 tokens per second with 3.90-second P50 latency, putting it below the latency profile of hosted US frontier APIs for real-time interactive applications.