PHYSIX FRONTIER EASTVOICE
2026.09.23 · Wednesday
Issue 045
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
50%
OpenAI API prices down roughly 50% with GPT-6 Sol and Luna
40%
Anthropic Claude Opus 5.5 costs roughly 40% below Opus 5
8 trillion
DeepSeek next target 8 trillion parameters, founder Liang Wenfeng
5 million
China active agents nearly 5 million, 2026 assessment report
Top Stories 3 · dual commentary ｜ Section Picks 25 · one-line takes
OpenAI Ships GPT-6 Sol and Luna, Anthropic's Opus 5.5, and Alibaba Confirms Qwen4 Is Training.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 23, trained along the same path as GPT-6 Astra, with API prices reported down roughly 50% and fewer errors on factual and coding tasks.
What it means for you｜API prices down roughly 50% and fewer errors on factual and coding tasks mean cheaper access for developers building on OpenAI models.
Editor's take｜Astra is barely a month old and OpenAI is already filling in the cheap seats. Price bands, not benchmark wins, are what close enterprise deals.
Anthropic launched Claude Opus 5.5 on September 23, the first model in its 5.5 family. The company claims performance comparable to Claude Fable 5.1 on most work at roughly 40% below Opus 5 running costs.
What it means for you｜Performance comparable to Claude Fable 5.1 at roughly 40% below Opus 5 running costs means lower bills for buyers of Anthropic models.
Editor's take｜A new family's first release anchors what buyers expect to pay for everything after it. Anthropic set that anchor deliberately low.
Reuters reported on September 22 that DeepSeek will appear before the UN Security Council this week to brief members on AI safety risks, as world leaders gather in New York for the General Assembly.
What it means for you｜DeepSeek briefing the UN Security Council reframes AI safety as a multilateral file rather than a US export-control file, shifting how AI risk is governed.
Editor's take｜A Chinese lab briefing the Security Council reframes AI safety as a multilateral file rather than a US export-control file. That shift matters more than any model shipped this week.
25 AI stories worth your time, each with a one-line take.
According to The Information, DeepSeek is training a model of roughly 2 trillion parameters, and founder Liang Wenfeng told investors recently that the next target is 8 trillion.
One-line take｜DeepSeek built its reputation on efficiency and is now entering the parameter race it once dismissed. Scale is a strategy only if you can serve it at a price.
At Yunqi on September 22, Alibaba said Qwen4 is being trained on a new architecture and that Qwen4.5 and Qwen5 will scale into the 5 to 10 trillion parameter range. Recursive self-improvement has already reached training, inference and chip-model co-design.
One-line take｜Naming the next three versions and their parameter band is a supply-chain signal, aimed at whoever is committing GPU budget for 2027.
Harvey, the legal AI company valued at $15.6 billion, is rethinking its reliance on OpenAI as inference and compute costs climb, and has been evaluating Chinese open-weight models, per a September 22 report.
What it means for you｜Harvey, valued at $15.6 billion, evaluating Chinese open-weight models as inference and compute costs climb gives US enterprises a cost-saving reference.
One-line take｜Cost pressure at the application layer is doing the work no procurement memo could. Chinese open weights just got their first serious US enterprise reference account.
Inspur unveiled the Yuanbrain SD200 Ultra supernode at AICC 2026, claiming a single unit can serve Kimi K3 at 2.8 trillion parameters, with token latency down to 5.85 milliseconds.
What it means for you｜A single unit serving Kimi K3 at 2.8 trillion parameters with token latency down to 5.85 milliseconds means frontier models no longer need a customer-assembled cluster.
One-line take｜The 5.85 ms figure matters more than the parameter count. Frontier Chinese models no longer need a cluster the customer has to assemble first.
Loongson released the first software version of its accelerated computing platform for in-house general-purpose GPUs, covering drivers, a programming environment, OpenCL 3.0, CUDA compatibility and AI inference.
What it means for you｜Loongson's software stack covering drivers, OpenCL 3.0, CUDA compatibility and AI inference gives users a supply-chain hedge for in-house GPUs.
One-line take｜Driver and compiler support is the part nobody budgets for and everybody gets blocked on. This is a supply-chain hedge, not a benchmark entry.
At Yunqi on September 22, Alibaba Cloud CTO Li Feifei laid out an "Agentic Cloud" strategy built on models, harnesses and context, with new products including AgentCore, Agent Sandbox and the next-generation CPFS storage system.
One-line take｜Treating harness and context as first-class cloud primitives is the clearest statement yet that Chinese cloud vendors expect agents, not people, to be their main API consumer.
Qwen said at Yunqi that it is accelerating work on a Personal Agent built on the agentic capabilities of the Qwen 3.8 series, with the stated goal that every user gets their own AI assistant.
One-line take｜Consumer agents live or die on the phone's default assistant slot. Alibaba is arguing for that slot through hardware rather than through the app store.
Alibaba's QwenWork released enterprise agent products on September 22, including enterprise context, digital employees and the QwenNote A2 agent hardware, pitched at agents that understand the business, collaborate with staff and can be trusted.
One-line take｜Enterprise context is the quiet layer of every agent deployment. Whoever owns it owns the switching cost.
Alibaba launched Qwen Intelligence, a full-stack AI phone solution, with Honor's Magic9 series, due September 28, as the first device to carry it. Alibaba calls it the industry's first system-level Agent Harness architecture.
What it means for you｜Honor's Magic9 series due September 28 as the first device to carry Qwen Intelligence puts the agent below the app layer, where a user cannot uninstall it.
One-line take｜A system-level harness puts the agent below the app layer, where a user cannot uninstall it. Google and Apple are fighting the same battle.
Tencent released Hy Image3.5 preview on September 22 with text-to-image, image-to-image and multi-turn creative dialogue, shipping into Yuanbao and ima. Tencent Video's WorkRally platform opened it to creators free for two weeks.
What it means for you｜Tencent Video's WorkRally platform opening Hy Image3.5 preview free for two weeks gives creators access to text-to-image, image-to-image and multi-turn creative dialogue.
One-line take｜Two free weeks for film and series production teams is how reference customers get made. Generated storyboards are the cheapest way into a studio's pipeline.
China Unicom released AICC3.0, moving its intelligent hotline from AI-assisted answers to autonomous task execution, so a call can finish a job instead of only explaining one.
What it means for you｜China Unicom's AICC3.0 moving its intelligent hotline to autonomous task execution means a call can finish a job instead of only explaining one.
One-line take｜Government and enterprise hotlines are agent deployments with a real success metric. Completion rate, not satisfaction score, is the number to watch.
The 2026 China AI Computing Power Development Assessment Report put active agents in China at nearly 5 million, projecting compound annual growth above 150% from 2026 to 2030 and a 300% expansion in the AI compute market.
What it means for you｜Active agents in China at nearly 5 million and a 300% expansion in the AI compute market project growth that utilities and chip vendors buy against.
One-line take｜Agent counts stay a vanity metric until someone reports how many survive past week one. The compute forecast is the number utilities and chip vendors actually buy against.
Alibaba showed the Qwen AI Glasses N1 and N1 Pro at Yunqi alongside clip-on AI earbuds. The glasses support eye tracking and iris payment, and go on sale October 13.
What it means for you｜Iris payment in the Qwen AI Glasses N1 and N1 Pro, on sale October 13, makes the glasses a payment terminal, changing who must approve them before shipping.
One-line take｜Iris payment makes the glasses a payment terminal, which changes who has to approve them before they ship. Payments compliance is now a wearable supply-chain problem.
Forbes' annual list of the 400 richest Americans put the entry bar at a record $4.4 billion. The number of billionaires under 40 rose from 4 to 10, and 7 of them built their fortunes on AI, four of them at Anthropic.
One-line take｜Four of those ten sit inside one company. Equity concentration at this scale shows where the returns of this cycle are landing.
Tesla managers and engineers have begun auditing Chinese suppliers including Sanhua Intelligent Controls, as the company works out how to mass-produce Optimus while Washington tightens restrictions on Chinese robotics.
One-line take｜Washington writes the export rules and Tesla shops the supply chain anyway. The humanoid bill of materials still runs through China.
Tesla placed orders for roughly 5,000 Optimus robots with suppliers and sent a team to Ningbo on September 16 to audit robot suppliers, according to supply-chain reports.
One-line take｜The distance from one robot to ten thousand is manufacturing engineering, and the car industry already has it. That is why Tesla keeps beating robotics startups on cost.
A public dispute between Mekamand founder Shao Tianlan and Galaxy General over alleged fabricated revenue, related-party transactions and a rushed listing has exposed how thin the commercial floor is under parts of China's embodied AI sector.
One-line take｜When founders start accusing each other of booking fake revenue in public, the funding environment has already turned. The audit trail will matter more than the demo reel from here.
Qiyuan Robotics, launched in late 2025, held a product event in Shanghai on September 20 with celebrity guests, and now has to justify its parent Shangwei New Materials' roughly ¥60 billion valuation with ¥20,000 consumer robots.
What it means for you｜Qiyuan Robotics pricing ¥20,000 consumer robots sets a volume expectation the company cannot meet yet, making its roughly ¥60 billion valuation a delivery schedule.
One-line take｜Pricing a robot like a home appliance sets a volume expectation the company cannot meet yet. The valuation is now a delivery schedule.
Nikkei reported that Japan plans to hold a robot competition aboard the International Space Station as early as 2027, with teams using AI to control robots and the stated goal of training the next generation of space engineers.
One-line take｜An orbital contest with a two-year runway is a talent pipeline dressed as a race. Robotics recruiting keeps starting earlier.
Mercedes-Benz agreed on September 22 to develop autonomous driving systems with UK startup Wayve, with at least one model carrying the system due within two years.
What it means for you｜Mercedes-Benz developing autonomous driving systems with Wayve, with at least one model due within two years, brings end-to-end autonomy to a luxury OEM platform.
One-line take｜Wayve's end-to-end stack just got a luxury OEM as its first real platform. For a company that bet against HD maps, German homologation is the proof point that counts.
Owners running FSD (Supervised) 14.3.9 report frequent alerts claiming the front camera is dirty and requiring a service visit, even when the windshield and camera housing are clean.
What it means for you｜FSD 14.3.9 alerts claiming the front camera is dirty on clean glass send owners to service centers for nothing.
One-line take｜A cleaning warning that fires on clean glass is a perception threshold problem, not a hardware one. False alarms at this rate send owners to service centers for nothing.
The Czech Republic formally approved Tesla's FSD (Supervised) system for use on public roads, adding another country to the European rollout list.
What it means for you｜Czech Republic approving Tesla's FSD (Supervised) for public roads adds another country to the European rollout list, making the next filing cheaper.
One-line take｜Country-by-country approvals are the only route into the European market. Each one makes the next filing cheaper.
A TMTPost analysis argues that while Robotaxi takes the headlines, fixed-route autonomous buses are the vehicles most likely to enter urban public transit at scale.
One-line take｜Fixed routes and depots remove the hardest part of autonomy, which is the long tail. Fleet economics get proven in the unglamorous segment first.
XPeng began rolling out XOS 6.3.0 for its second-generation VLA system on September 22, describing a move from 3D spatial recognition to 4D spatiotemporal understanding through stronger temporal memory, trajectory prediction and decision response.
What it means for you｜XPeng's XOS 6.3.0 rolling out second-generation VLA from 3D spatial recognition to 4D spatiotemporal understanding brings stronger temporal memory and trajectory prediction to a car.
One-line take｜Adding time to the representation is the same problem world-model teams are working on, shipped to a car first. Driving stays the fastest route for spatiotemporal models into production.
At Yunqi, Alibaba confirmed Qwen4 is in training and that its next-generation video model is also underway, alongside progress on a native omni-modal generation model the company expects within three years.
One-line take｜Video generation is where Chinese labs are placing their world-model bets, because the training data arrives free with the product. Whether that also buys physical understanding is still open.
📖 The full EastVoice issue on the Chinese forum
📈 Physix Frontier · Alpha — the investor's read
Headline teardowns · curated signals · market moves. Five minutes to find the alpha of the AI era →