---

Three threads today. One in the courtroom: a US appeals court upheld, 2 to 1, the Pentagon's supply chain risk designation against Anthropic, keeping the company's models out of US military systems. One on the books: Anthropic's seven founders want 50.1% of the voting power before going public; Microsoft merged chat, coding, and assistant into one Copilot super app. And one on method: TMTPost tore down the freshly released Opus 5.5 and GPT-6 Luna, and found that the two models' approach to saving compute (the machine resources needed to run AI) had already appeared in Chinese teams' papers and announcements. A-shares are closed Saturday; quotes use the September 25 Friday closing prices.

---

Six, Macro and Market Data

Two new US models use the savings tricks Chinese teams proposed first

TMTPost tore down two freshly released US models: Anthropic's Opus 5.5 and OpenAI's GPT-6 Luna. The author found that their approach to saving compute (the machine resources needed to run AI) had long since appeared in Chinese teams' papers and announcements. Down to the numbers. Opus 5.5, at its highest setting, outputs an average of 119,000 tokens per question (the smallest unit of text a model reads and writes), of which 84,000 are the thinking process — it takes the route of thinking more on the spot to buy accuracy. GPT-6 Luna goes the other way: $0.1 per million tokens input, $0.5 output, half the price of the previous generation; cache hits (storing already-computed repeated content to fetch directly next time) cost $0.01, 90% cheaper. DeepSeek launched the same approach on August 2, with cache hits at 0.1 yuan per million tokens (the smallest unit of text a model reads and writes), also 90% cheaper. If you're a developer paying by call volume, or buying AI tools for a team, watch cache pricing over the next few months. Put system prompts and long documents — the stuff read every time — at the very front; once they hit the cache, the unit price for the same work drops by ninety percent. When picking a model, add one question: what price does the repeated part get charged at.

Changan and CATL launched a project for sodium-battery swap, targeting extreme cold

Per ITHome on September 25, from Changan Automobile's investor relations activity record published September 24. The company made phased progress on sodium-ion batteries, breaking through single-point technology for high-power electrodes, used to improve range under extreme temperatures. It also set up a project with CATL for sodium new battery-swap technology, the two jointly developing a new-system battery. Sodium batteries' strength is cold resistance. CATL released its sodium new battery in April 2024, rated for use from minus 40°C to plus 70°C; installed in a car with a 2.95-meter wheelbase, the hybrid version's pure-electric range is claimed to exceed 200 kilometers, and the pure-electric version over 500 kilometers. Battery swap means taking the battery off the car and swapping in another, a matter of minutes, no waiting in line at a charging pile. People who use their cars year-round below minus twenty degrees can watch the rollout timetable for such announcements. When buying a car, ask about measured low-temperature range — more useful than looking at the rated mileage. Readers driving for ride-hailing can wait for swap stations to spread before considering it; what battery swap saves is the time of charging once a day.

Microsoft stuffed chat, coding, and assistant into one interface; Copilot's new positioning is for work

Microsoft released a new version of Copilot on the evening of September 25. The Information's report says this time it put chat, coding, and agents (AI that can take a few steps on its own) into the same interface, directly targeting Anthropic's spot in the enterprise market. Microsoft positions it as AI born for work, and even compared it to Office's impact on productivity back in the day. There's a backdrop to this move: in the enterprise market, Anthropic's models grabbed a lot of coding orders this year, and Microsoft needs to bundle its several capabilities into one entry point. Microsoft stock rose 3.66% that day, the brightest of the six US AI names. Readers using Copilot at work will next see the interfaces merge, with the previously separate chat, coding, and automation entries gathered in one place. When picking tools, try one standard: do the same task once in chat mode and once in assistant mode, and see which makes you redo less work. Individual users don't need to rush to switch; the enterprise version moves first.

Appeals court upholds Pentagon ban, Anthropic's models can't enter US military systems

The US Court of Appeals for the District of Columbia Circuit upheld, 2 to 1, the Pentagon's supply chain risk designation against Anthropic on September 25. Under this designation, the US military cannot use Anthropic's Claude models, and defense contractors cannot use them on government projects either. The starting point was the $200 million contract in July 2025, and the negotiations that September about deploying Claude on military AI platforms. The Pentagon wanted all lawful uses allowed; Anthropic wanted two bottom lines: not for fully autonomous weapons, not for domestic mass surveillance. After talks broke down, Defense Secretary Hegseth accused the company of wanting veto power over US military operational decisions. A federal judge in San Francisco already ruled one of the designations unlawful last month; this time the other one was upheld. AI companies doing government and enterprise business can remember the consequence of this case: writing use restrictions into a contract may be deemed a supply chain risk. When drafting proposals for government projects, prepare the use clauses and a backup supplier together — don't let one line's model become the only option.

• Leiphone counted 11 papers Liang Wenfeng has co-authored in the past three years. He doesn't appear in the flagship model reports with hundreds of authors, like V3.2 and V4; he only signs on underlying technologies like MoE routing (splitting a large model into many small experts, called on demand), attention mechanisms, mHC, DSpark, and DSec. The latest is DSec, uploaded September 19, an automated sandbox: it can run 3 million isolated environments a day, sustain 380,000 concurrent instances, with a single production unit spanning 160 CPU nodes, and it includes anti-cheating mechanisms. People building AI products can note this sandbox's approach — when a bunch of assistants train together, environments and isolation are the bottleneck. Users won't see a change for now, but it determines whether the next generation of models can work reliably.
• TechCrunch, citing The Information: Anthropic is asking shareholders to approve a new equity structure giving CEO Dario Amodei and six co-founders special shares totaling 50.1% of voting power, on the condition that at least three of them maintain a minimum stake. Such special shares carry no extra economic benefit. The seven each hold only about 2% of the shares, and have pledged to donate eighty percent of their wealth. The company was valued at $965 billion in May, with the secondary market recently quoting $1.5 trillion. People watching this company's IPO can remember one thing: voting power and ownership percentage are two different things — asking who's in charge gets you closer to the truth than staring at the valuation number.
• Electrek reported on Geely's fifth-generation smart charging station launched in Ningbo, with peak power of 2250 kilowatts. In testing, the batteries of the Lynk 10 and Zeekr 001 went from 10% to 70% in 4 minutes 30 seconds, and to 97% in 8 minutes 40 seconds. The cost of fast charging is that lithium ions pile up in a layer on the anode, and the battery slowly loses capacity. Geely's approach is to pulse micro-currents into the cells, reactivating the piled-up lithium ions, paired with AI-managed charging, extending cycle life by up to 20%. Temperature is managed by an energy model co-developed by Geely and StepFun, which can predict battery temperature 30 seconds ahead, keeping the average temperature during charging under 55°C. EV drivers can watch one thing: home slow-charging strategy is also being upgraded — fast charging on long trips, slow charging at home for repair, is the usage this scheme is designed for.
• MIT Technology Review dug this project out of the US Department of Defense's FY2027 budget request: $30.3 million over five years for a lie detection system called Polygraph+. It combines an AI scoring algorithm with a standoff sensing technology, reading physiological signals without wiring up the subject, for onboarding screening and internal leak investigations, run by DCSA, which handles background checks. The backdrop is a September New York Times report that about 50 Joint Staff officers were polygraphed. Kotsoglou, a scholar studying lie detection at Northumbria University in the UK, said this is forcing something complex into a measurable number. Workers can remember one thing: when AI scoring enters HR processes, asking what the appeal channel is is more practical than arguing about whether it's accurate.
• An analysis posted on Substack puts two things side by side. In classic cloud, everyone is a customer, with AWS, Azure, and GCP as the three landlords; in AI it's the reverse — everyone wants to be the landlord, and the real customers number only about eight: a few model labs, plus companies doing inference services with open-source models. The author's example is CoreWeave, which got more than half its revenue from Microsoft at IPO. The most effective sales pitch right now is one line: we have GPUs. People watching this thread can note that once GPUs are no longer scarce, buyers will sit down to negotiate terms, and the bargaining initiative returns to the buyer's side.
• A European discussion paper — A discussion paper on the EU AI Alliance platform raises a question: if an attacker controls a company's internal AI assistant connected to email, finance, R&D, and HR, can it use its legitimate permissions to piece together the company's present and future strategy to near 85% accuracy, even if no single document fully spells out that strategy. What the author presses on is technical control: what means exist to stop this reconstructed strategy from being stored, transmitted, or turned directly into action. The piece ends by listing several IETF drafts. A concrete action for enterprise IT and compliance readers: list the systems the AI assistant can see, mark which few connected together would be enough to piece together the strategy map, and start by reining in permissions there.
• Changan says the auto industry has entered a stock-game, with profit margins falling to historic lows — Changan says the auto industry has entered a stock-game, with profit margins falling to historic lows. Per ITHome on September 25, sourced from Changan's September 24 investor relations record. The same material has two more points: the company has full coverage of batteries, motors, and electronic controls; it secured China's first official L3 autonomous driving license plate last December, and this year was approved for an L4 self-developed Robotaxi test license. L3 and L4 are autonomous driving levels — the higher the level, the fewer times a human needs to take over. Car watchers can note one line: when the price war stalls, manufacturers shift money to smart driving, and mid-trim versions of new cars in the second half may carry an extra lidar.
• A Huxiu piece starts from the 800 yuan extra on your iPhone and follows the price upstream — A Huxiu piece starts from the 800 yuan extra on your iPhone and follows the price upstream: first to memory and HBM (the fast storage AI chips need), and finally to AI chips costing tens of thousands of dollars apiece. The piece names several domestic manufacturers: Huawei, Xiaomi, Honor, OPPO, vivo. The part relevant to upgrading your phone is here: phones and AI compete for the same batch of memory capacity, and part of the extra cost is paying the bill for this round of price hikes. People planning to upgrade can wait for new model prices to be announced before deciding which tier to buy.
• NOEMA published a long piece suggesting a different way to understand AI: like a dragonfly, using compound eyes to look in nine directions at once. The author, Anthea Roberts, is a professor of international law and global governance, and she groups the nine current claims about AI into opposing pairs — NOEMA published a long piece suggesting a different way to understand AI: like a dragonfly, using compound eyes to look in nine directions at once. The author, Anthea Roberts, is a professor of international law and global governance, and she groups the nine current claims about AI into opposing pairs — builders and the displaced, geopolitical hawks and power critics, information disruptors and truth guardians, with the other three (environmental, safety, humanistic) having no corresponding benefit side. Her conclusion: measuring AI with only one ruler amounts to deciding on certain losses for others, often environmental costs and cognitive decline that don't enter GDP. One point relevant to ordinary people: when you see one-sided AI commentary, first ask what it left out of the calculation.
• A tech blog says a single AI call in production is actually a hidden program — A tech blog says a single AI call in production is actually a hidden program. The author at Seldon lays out this kind of work: having a model read annual reports and extract customer-supplier relationships sounds like general reasoning (the model writing out its thinking then answering), but unpacked it's a fixed sequence of extraction, normalization, entity alignment, relation judgment, and format validation. His advice is to first name these operations, then recover the program from real traffic, and stop handing to the model the parts that can run on cheap machines. Readers building AI applications can try once: unpack your most expensive call and see how many steps in it don't need the model at all.
• A long piece on the advertising business — A long piece on the advertising business: after AI made images and short videos nearly free, the flood of low-quality content thinned the value of micro-video ads. Author Daniel de Boulay's judgment is that money is moving the other way, toward narrative works people will pay to watch repeatedly. His example is Saint Laurent co-funding Almodóvar's new film. Readers in branding and content can note one action: split the budget in two — one part buys exposure, one part buys works that can still be verified as genuinely filmed ten years from now.
• Someone added a third category to the term "AI psychosis": the kind that treats AI as an acquaintance. Author Jeff splits the popular usage into three: genuine psychosis, high-volume but worthless output, and the parasitic-social type he writes about — Someone added a third category to the term "AI psychosis": the kind that treats AI as an acquaintance. Author Jeff splits the popular usage into three: genuine psychosis, high-volume but worthless output, and the parasitic-social type he writes about — through repeated conversations with AI, a person develops a one-way relationship with it like with a streamer or a celebrity. He reminds us the brain is naturally unable to tell two-way from one-way relationships. People who chat with AI often can self-check one thing: take your most frequent topic and talk it over with a real person once, and see whether you still want to speak.
• GitHub added a canvas feature to Copilot — GitHub added a canvas feature to Copilot: describe the interface you want in one plain sentence, and the assistant generates the kanban board, form, or checklist directly, and you and it can edit it together. The official guidance for beginners is to first state three things clearly: which process this interface supports, what you can do in it, and what the assistant can do. The example in the blog is a release notes board. Readers coding with Copilot can try once, handing it that table you remake every week in one sentence.
• An AI experiment popped up on Hacker News — An AI experiment popped up on Hacker News: a public map where anyone can claim a plot for free, and then a team of AI assistants builds an annotated simulated scene on your plot. The project is called The District. If you want to see what a bunch of AIs working together looks like, spend ten minutes claiming a plot this weekend — more intuitive than reading papers.
• Bloomberg reported that Google DeepMind saw a wave of departures, and VCs are chasing this group to invest in AI's next big direction. Readers following AI startups and funding can watch new fund announcements over the next few months; this group's names will likely appear on the list.
• Bloomberg reported that Trump demanded OpenAI and Anthropic not hand their models to UK institutions. People doing overseas business can note the practical impact: whether a model can be evaluated for a given country starts being decided by political stance, no longer just price and capability. When companies pick model suppliers, it's worth asking one more question about the host country's policy restrictions.

---

Today's Market Quick Look (in-house quotes, global 24 hours (US stocks 9/25 (Friday) close + A-shares 9/25 (Friday) close))

Four up, two down. Microsoft rose 3.66%, the brightest of the six, the day it merged chat, coding, and assistant capabilities into the same Copilot interface. Alphabet, Nvidia, and Amazon all rose less than half a point. The two decliners were Tesla and Meta; Meta fell 3.33%, giving back gains from last week's developer conference.

All six down. The three doing optical modules and server assembly fell hardest, all around 3%. Domestic chips and office software fell one to two points. Today is Saturday, the market is closed, and these quotes use the September 25 Friday closing prices. There was no volume expansion; A-shares are still on the sidelines on AI pricing.

---