China Drafts AI Export Controls as Anthropic Pitches a Federal Safety Bill and Claude Redesigns Quantum Circuits
Rules for frontier AI arrived from two capitals on the same Friday. China’s commerce and science ministries proposed export controls on 23 AI technologies, and Anthropic confirmed it is circulating a draft federal bill that would let a judge halt a dangerous model. Anthropic also showed Claude Mythos 5 redesigning quantum circuits a day after Google demonstrated magic state cultivation on Willow. OpenAI explained how its own engineers maintain code with GPT-6 models and added parallel agents to GPT-6 Pro, Reflection put Beam’s weights on Hugging Face, and Liquid AI shipped a calibrated decision model.
China Proposes Export Controls on 23 AI Technologies, With Open Source Carved Out
China’s Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48 of 2026 on October 16. The Ministry of Commerce announcement on revising China’s export-control technology catalog proposes adding 40 technologies in AI, biomedicine, and photovoltaics, 23 of them in AI covering large models, agents, computing clusters, and information processing. Technologies already public through open source or published papers are excluded. Comments run until November 15, and until the revision takes effect, exporters must register covered technologies before exporting them, including exports already underway.
The Global Times report on the draft revised export catalog says 14 additions are prohibited and 26 restricted. The draft would ban exports of core technologies for large models trained on national key sensitive data and of AI agents serving key national information infrastructure, national defense, and core government affairs, and would restrict 2.5D and 3D AI chip design technologies. In a Ministry of Commerce spokesperson Q&A on the catalog revision, published in English from CCTV, the ministry said automatic licensing would cover open-source AI exports and Chinese companies building self-operated overseas computing centers. Overseas localized deployment of a restricted model would count as an export, but ordinary services would not.
"For situations where enterprises use restricted large models to provide external API calls, APP/web page interactions, SaaS applications, local end-side lightweight deployment, etc., they do not fall within the scope of large model technology export management." — Ministry of Commerce spokesperson, via CCTV
For multinationals, the key line falls between serving a Chinese model and transferring it. API access, SaaS, and lightweight on-device deployment stay outside the draft’s scope, but overseas localized deployment of a restricted model does not. Companies that license Chinese model technology or co-develop agents with Chinese partners should map those deals against the draft list and file comments before November 15.
Anthropic Circulates a Draft Federal Frontier-AI Safety Bill
Anthropic confirmed on October 16 that it is circulating a draft federal AI bill among U.S. lawmakers, after Politico first reported it. Under Reuters’ summary of Politico’s report on Anthropic’s draft bill, the government could block a model’s release or use only for a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people, stemming from a biological weapon, the model evading its developer’s controls, or a cyberattack. Evaluators registered with the Commerce Department would periodically assess labs, could be paid as long as payment is not tied to the outcome, and would have 72 hours to report “imminent enumerated risk.” Commerce could ask a federal judge in the District of Columbia to suspend a model, and no other agency could.
An Anthropic spokesperson told Bloomberg, in the Bloomberg report on Anthropic’s AI safety bill carried by Moneycontrol, that the bill would require labs to test models, share findings publicly, and allow routine independent review.
"This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June" — Anthropic spokesperson, via Bloomberg
The AFP report on Anthropic’s proposed AI law in The Hindu says the bill would override certain state laws, such as California’s. Congress is out of session until after next month’s midterms. Charlie Bullock of the Institute for Law & AI called it “pretty reasonable,” but “weaker” than the FRONTIER Act, a competing House bill.
Treat this as a preview of what federal compliance might look like, not as law. If something like it advances, frontier developers would need evaluator contracts with outcome-independent fees and a 72-hour reporting path, while enterprises building on those models would mostly inherit disclosure documents. Watch the preemption language closely, because companies under California rules could see obligations replaced rather than added.
Claude Mythos 5 Redesigns Quantum Circuits as Google Cultivates Magic States on Willow
Anthropic’s research post on designing more efficient quantum circuits with Claude, published October 16 from its STEM Fellows Program, says Claude Mythos 5 received plain-language specifications for eight circuits from chemistry, optimization, physics simulation, and number theory, plus an automated evaluator that scored each design on qubits multiplied by gates. Working in a closed loop with no human input between rounds, it found designs that used both fewer qubits and fewer gates than the best previously published ones.
"Working largely on its own, Claude found designs that improve on the best published circuits, in some cases by an order of magnitude or more." — Anthropic
For elliptic-curve point addition, the core step in quantum algorithms for the elliptic-curve discrete logarithm problem, a 53-hour run produced the first reversible divsteps modular inversion with no stored branch history, verified in Lean. One resulting design uses 12% more qubits but 17 times fewer gates than the smallest previous one, and a redesigned decoder for decoded quantum interferometry cut gates 12-fold versus Google’s prior construction. Anthropic noted that Claude was conservative early on, so researchers required parallel, distinct approaches.
A day earlier, Google’s post on magic state cultivation on its Willow processor reported about 40 times fewer errors than un-cultivated states and T-states with a record 99.99% logical fidelity, published in Science. Google says some estimates put up to 95% of a quantum computer’s capacity into making magic states.
"Of course, we still need to demonstrate magic state cultivation at a larger scale to make it realistic for practical quantum computing, and plenty of technical challenges remain." — Google Quantum AI
Both results shrink the same budget, the qubits and gates a useful quantum algorithm needs. Anthropic’s method also travels well beyond quantum, since it works wherever a design can be written as code, checked automatically, and scored by cost. R&D teams in chip layout, compilers, or logistics should copy that verifier-first loop. Security teams should keep post-quantum migration plans on schedule.
OpenAI Shows How GPT-6 Models Maintain Its Own Code and Adds Parallel Agents to GPT-6 Pro
OpenAI’s developer blog on maintaining software with the GPT-6 family, by an engineer on its Core Experimentation team, describes GPT-6 Astra in Codex working from a vague dead-code prompt, checking pull requests, Slack threads, and production profiling before proposing deletions. Five engineers, part-time over two weeks, merged more than 1,400 pull requests and deleted over 800,000 lines of code. Astra also flagged a “Permanent” feature flag causing CPU fan-out for human review instead of deleting it; removing it saved an estimated 0.42% of CPU in the main affected service. GPT-6.1 Sol Ultra produced 92% as many stale-experiment findings as GPT-6 Astra Ultra at 14% of the API cost, $15.12 versus $109.06.
"However, some proposals added more complexity than their expected benefit justified, so prioritization and measurement remained an important part of our workflow." — OpenAI Core Experimentation team
The OpenAI Help Center ChatGPT release notes on the GPT-6 Pro update say GPT-6 Pro in the Chat tab now responds faster, uses plugins and tools better, and runs multiple agents on parts of a task in parallel, each reviewable when done. OpenAI’s announcement post on parallel agents in GPT-6 Astra Pro confirms availability for Pro users.
The 92 percent at 14 percent comparison is the number engineering leaders should keep. Pilot maintenance agents on the cheaper model with a narrow scope, collect what it misses, and patch prompts for those patterns, as OpenAI did with dependent flags. Keep a human gate for anything marked permanent or tied to incident response, and measure each cleanup’s benefit before merging.
Reflection Puts Beam’s 501B Open Weights on Hugging Face Under Apache 2.0
Reflection previewed Beam on October 5 and promised the weights later in the month; they arrived on October 16. The Hugging Face launch post for Reflection’s Beam describes a text-only mixture-of-experts model with 501B parameters, 23B active per token, a 1M-token context window, and an Apache 2.0 license, in BF16, FP8, and NVFP4 with day-0 support in transformers, SGLang, and vLLM.
"All 256 experts stay in memory, so deployment needs the memory of a 501B model but runs at the speed of a 23B one." — Hugging Face
Reflection’s developer guide to downloading Beam’s weights lists four NVIDIA B300 GPUs for self-serving the 273 GiB NVFP4 checkpoint. Benchmarks Reflection supplied for the Hugging Face post put the BF16 checkpoint at 81.6% on SWE-bench Verified.
Apache 2.0 removes the license review that slows adoption of other large open models, so Beam belongs on any shortlist for sovereign or air-gapped coding agents. The real constraint is hardware, since Reflection’s serving guide calls for four B300 GPUs even for the NVFP4 checkpoint. Rerun Hugging Face’s published vibe-eval tasks against your own work, and treat the benchmark table as vendor-reported.
Liquid AI Ships d1-ultra, a Decision Model Trained for Calibrated Answers
Liquid AI’s announcement of the d1-ultra decision model, released October 15, describes a multimodal model that reads text, images, or both plus questions in one forward pass and returns a probability for each answer without generating tokens. It costs $0.042 per million input tokens with no output charge. Liquid says it beats Jev 1.14 across all tasks in 112 evaluations and leads other decision models on the Artificial Analysis Decision Index by at least 8 points.
"Accuracy is not enough: decision models also output probabilities that must be properly calibrated." — Liquid AI
Decision models fit triage, routing, moderation, and visual inspection steps that currently spend generative tokens on yes-or-no answers. Calibration is the claim to test. Set automation thresholds from the returned probabilities and send the uncertain middle band to people. Measure calibration on your own labeled cases before relying on any vendor-reported index, including this one.
Today’s thread is where the lines get drawn. Beijing and an American lab both proposed rules on the same day, each trying to separate a narrow set of truly dangerous capabilities from everyday use. Meanwhile capability kept spreading: Claude designing better quantum circuits, a cheaper OpenAI model nearly matching its flagship on real maintenance work, Beam’s weights under a permissive license, and calibrated decision models priced per input token. Watch the boundary definitions, because they will decide which of these tools your teams can use, export, or deploy.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →