▒▓███▓▒ ▒▓███▓▒ ▒▓███▓▒
▒▓▓ ▓▓▒ ▒▓▓ ▓▓▒ ▒▓▓ ▓▓▒
░▒ ▒░░▒ ▒░░▒ ▒░
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
──────────────────────────────────────────
No. 73 · Wednesday, 23 September 2026Wednesday, and ten days after Anthropic's chief executive asked the industry to slow down, Anthropic and OpenAI shipped cheaper models on Tuesday about 90 minutes apart; Anthropic says its new Opus beats its own larger Fable on several benchmarks, and METR, which tested it before launch, calls it a step, not a jump. The Information reports that Beijing is investigating DeepSeek and Moonshot, two of the labs Anthropic says fed their users' conversations into Claude. At the UN, Trump said AI would now be called super intelligence in US documents, and the hotline to Beijing we flagged yesterday is an open line between two officials. Xi Jinping meets Trump at the White House this week.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Anthropic released Claude Opus 5.5, which it says beats its larger Fable 5.1 on several of its own benchmarks at a lower price
Anthropic put Opus 5.5 on sale on Tuesday at $4 per million input tokens and $20 per million output, 20% below Opus 5, and says typical workloads cost about 40% less because the model also uses fewer tokens to finish a job. On the company's own tests it beats Fable 5.1, Anthropic's bigger and pricier model, on agentic coding, knowledge work and computer use, and Anthropic called it "the strongest-performing model we've tested to date" on its automated behavioral audit; Sonnet 5.5 and Haiku 5.5 follow in the coming weeks. It is Anthropic's first model release since its chief executive, Dario Amodei, asked the industry to pace the frontier, which we covered on September 14, and it arrived a day after our Monday edition reported, citing Reuters, that the company was considering a new model to answer OpenAI. Amodei's essay asked the labs to slow the pace of capability gains and said pacing does not mean halting model training or technical progress; METR, a nonprofit that tested the model before launch, calls this one an incremental step past Fable 5.1 on its AI-research tests rather than a jump (more in PAPERS, below), which makes it a cheaper way to buy roughly what Fable 5.1 already delivered (Anthropic · TechCrunch).
OpenAI launched GPT-6 Sol and Luna at half the price of the models they replace
About 90 minutes later OpenAI released two cheaper members of the GPT-6 family, which it says carry GPT-6 Astra's gains in coding, factuality, computer use and alignment at a much lower cost. Sol costs $2 per million input tokens and $10 per million output, half of GPT-5.6 Sol's promotional price; Luna, built for high-volume work, costs $0.10 and $0.50. Opus 5.5's new price matches GPT-5.6 Sol's promotional price, and GPT-6 Sol costs half of it. OpenAI also published principles for outside safety assessments of frontier models on the same day (more in AI, below) (OpenAI · VentureBeat · TechCrunch).
China's internet regulator is reported to be investigating DeepSeek and Moonshot after Anthropic said they fed their users' conversations into Claude
The Cyberspace Administration of China, the country's internet regulator, first summoned all seven Chinese labs named in Anthropic's September threat report and then narrowed its inquiry to DeepSeek and Moonshot AI, The Information reported on Tuesday, citing people with knowledge of the matter. Anthropic's 154-page report alleges that the two companies passed their own users' requests to Claude, including, in DeepSeek's case, requests from engineers building a police surveillance system; the regulator is examining whether sensitive police, military and state-linked corporate data reached an American model. Both labs have also been invited to make statements at today's UN Security Council briefing on AI, one source told Reuters; OpenAI's chief executive, Sam Altman, plans to address the session; Amodei is set to do so remotely (The Information · Gizmodo · Security Council Report · Quartz · ABC News).
Trump told the UN he rejects any global scheme to control AI, and the incident hotline offered to China turns out to be a channel between two officials (update)
Speaking to the General Assembly on Tuesday, the president rejected international constraints on the technology and said that from now on US government documents would call it super intelligence, because, he said, the word artificial makes it sound fake. The AI incident hotline with Beijing that we flagged yesterday is, according to four people familiar with it who spoke to Politico, an open line of communication between Treasury Secretary Scott Bessent and Vice Premier He Lifeng, and Chinese officials have not publicly agreed to it. Politico notes that, for now, it is not a formal system for the two governments to report or jointly respond to AI threats, ahead of Xi's White House meeting with Trump this week (The Hill · The Washington Post · Politico).
Alibaba unveiled its most powerful AI chip and set a target of 20 gigawatts of data-center capacity by 2032
At its Apsara conference on Tuesday, Alibaba showed the Zhenwu V900, from its T-Head chip unit, which it says triples the performance of its predecessor and describes as China's most powerful AI chip; mass production starts in the first quarter of 2027. The company also said Qwen 4 is already in training, with later models of up to 10 trillion parameters planned, and set a target of 20 gigawatts of data-center capacity by 2032 (Quartz · KrASIA).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░OpenAI published principles for outside safety testing, and was reported to have neared a testing deal with Anthropic. The framework OpenAI released on Tuesday covers four areas where third parties should assess frontier labs: safety cases, critical safeguards, dangerous-capability evaluations such as biology and cyber, and investigations of critical misalignment incidents, with principles such as agreeing in advance on what is being assessed and giving the lab, where appropriate, time to fix problems before publication. The framework came a day after the policy paper we covered yesterday, in which OpenAI asked Washington to lead on shared rules for testing AI models. The Information reported on Monday, September 21, that OpenAI and Anthropic got as far as a legally binding draft to stress-test each other's commercially available models, without keeping each other's data, and that it is unclear whether anything was finalized (OpenAI · The Information).
Meta says its Muse agent was heavily inspired by OpenClaw, after users found matching files down to one called SOUL.md (update). Nat Friedman, who runs product at Meta Superintelligence Labs, said on X that Muse was built from scratch but modeled as a product on OpenClaw, the open-source agent by Peter Steinberger, after users found identical workspace filenames with nearly identical contents. Friedman said Meta's team thought Steinberger got those things exactly right. The admission followed Amazon's move to block Muse from its store, which we reported yesterday, and came soon after Muse reached No. 1 on the US App Store (TechCrunch).
Cisco's new scanner found malware that takes orders from a committee of four chatbots. Cisco Talos, the company's threat-research unit, released CAIRN, an open-source tool that finds AI-driven malware by looking for prompt templates and model API endpoints without running the code. Its headline find, an implant called CLOSEDQUORUM, asks DeepSeek, Qwen, Mistral and Gemini what to do next and does whatever gets the most votes: steal credentials, inject, persist or move. Talos has not confirmed that it has been used against anyone, and the public build ships with dummy API keys (WIRED · Runtime Wire).
░░▒▒▓▓ TECH ▓▓▒▒░░ShinyHunters says it breached the FBI and took data on agents and job applicants. The extortion group posted the claim on its leak site on Tuesday and says it got in through an Oracle PeopleSoft server used for recruiting, then moved into an Amazon-hosted government system; 404 Media, which first reported the breach, verified part of a sample of agents' names, home addresses and phone numbers. The group says the hack is not about money: it wants the FBI to withdraw a May report on ShinyHunters' methods, the one that advised victims not to pay. The FBI said it is aware of claims of unauthorized activity affecting FBIjobs.gov and is investigating (TechCrunch · US News · CyberScoop).
Microsoft cut fewer than 600 more jobs and handed the next Halo game to Activision. The cuts announced on Tuesday include 268 roles at Xbox Game Studios, Halo Studios among them, and others in cloud, AI, marketing and research, about 300 of them in Washington state; Microsoft put the total at fewer than 600 globally. A new team at Activision, the Call of Duty publisher Microsoft owns, will build the next Halo, a franchise that began as Xbox's own flagship (GeekWire · Bloomberg).
░░▒▒▓▓ POLICY ▓▓▒▒░░California now has what its governor calls the most comprehensive data-center laws in the nation. Governor Gavin Newsom signed seven bills on Monday, September 21, that require proposed data centers to tell local governments and water suppliers about their water use, make data centers report their annual water use and try to keep the cost of grid upgrades for large data centers off other customers' bills. He signed them in the face of growing public backlash against data-center construction, CalMatters reports, a year after he vetoed a water-disclosure bill (Governor of California · CalMatters · Bloomberg Government).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Data-center IPOs are slowing. SB Energy, the SoftBank-backed energy developer, is not expected to list before mid- to late October after struggling to find buyers at the $50 billion-plus valuation it wanted, according to a report on Monday, September 21; a person familiar with the matter told Investing.com on Tuesday that the offering is advancing according to plan and that SB Energy never set a September launch window. It has yet to open a data center. Holtec, the nuclear company, paused its offering indefinitely, citing uncertainty around data-center development, and the power-rental firm Aggreko has slowed its own. We covered SB Energy's Nasdaq filing on September 2 and rising yields on data-center bonds yesterday (Investing.com · Business Standard · Investing.com, via Yahoo Finance).
AI providers now take 8% of software spending at the companies Zip tracks, up from 1.4% a year ago. Data from Zip, a procurement-software company, shows spending on OpenAI, Anthropic, the coding-tool maker Cursor, the customer-service AI company Sierra and similar providers rising sharply while overall software budgets grew a median 13%, The Information reported on Tuesday. Workday, HubSpot, Microsoft and Amazon are responding with discounts and free trials on their own AI tools while moving to usage-based pricing; one customer saw its monthly GitHub Copilot bill rise from $20,000 to $260,000 after Microsoft made that switch (The Information).
Snorkel AI, which helps AI labs build training data, nearly tripled its valuation to $3.5 billion. The seven-year-old company raised a $350 million Series E led by Insight Partners and S32, up from a $1.3 billion valuation last year, and says its annualized revenue run rate passed $375 million and that it works with frontier labs and US government agencies (Snorkel · TechCrunch).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The United States, Denmark and Greenland signed their Arctic defense agreement (update). The agreement, signed on Tuesday on the sidelines of the UN General Assembly after Trump announced the deal on Friday, which we covered on Monday, lets Washington expand Pituffik Space Base, set up bases at Narsarsuaq and Mestersvig and fly over and land anywhere in Greenland, and bars non-NATO countries from a persistent military presence without all three governments' approval. It takes effect only once Denmark and Greenland complete their parliamentary procedures (Al Jazeera · US News · The White House).
Britain will refuel Saudi jets defending against Houthi attacks. Prime Minister Andy Burnham said on his way to New York that one RAF Voyager tanker will support Saudi aircraft on defensive missions against drones and missiles for a matter of weeks, not strikes on the Houthis, and that the arrangement will be kept under review (Al Jazeera · The Defense Post).
The EU agreed to drop the billionaires Alisher Usmanov and Mikhail Fridman from its Russia sanctions list. EU envoys agreed on Tuesday to delist the two billionaires in exchange for renewing sanctions on the remaining roughly 3,000 people and entities for three years, after France and Luxembourg pushed for the removals and Latvia, the last holdout, abstained. Ukraine's foreign minister, Andrii Sybiha, called the decision "shameful and unjustified" (Ukrinform · The Moscow Times · Meduza).
░░▒▒▓▓ PAPERS ▓▓▒▒░░METR — Claude Opus 5.5 is an incremental step over Fable 5.1, not a jump toward automating AI research. METR, a nonprofit that tests whether frontier models could do AI research on their own, published its pre-release evaluation on Tuesday after ten business days of API access. It found Opus 5.5 an incremental improvement on Fable 5.1 on its quantitative tasks, with no sign of a large gain in the foresight and research judgment that METR expects full automation of AI research to require, though it calls that expectation highly uncertain. The evaluation was unpaid; METR drafted its summary for Anthropic's system card, Anthropic had the opportunity to edit it, and METR signed off (METR).
░░▒▒▓▓ REPORTS ▓▓▒▒░░The Economist/YouGov — Many Americans expect big benefits from AI but concerns are rising quickly. In the poll of September 18 to 21, 42% of Americans called AI's risks very serious, up from 35% three weeks earlier, and much of the rise came from Democrats, 53% of whom now say so against 39% before; Republicans barely moved, at 29%. Still, 47% say AI will have many positive benefits, against 31% who say it won't. On September 15 we covered a separate YouGov survey in which worry that AI could end humanity rose among liberals while conservatives held steady (YouGov).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Paramount agreed to set up a board of five veteran journalists to settle editorial disputes at CNN and CBS News. Paramount's settlement with 12 state attorneys general, announced on Monday, September 21, paves the way for its purchase of Warner Bros. Discovery, pending a judge's approval, and would set up a board of five journalists, active or retired, with at least ten years in the trade each, to write guiding principles for accuracy, independence and fairness and to settle disputes between newsroom staff and management over them. Colorado and Washington are excluded from the board provision (Deadline · Variety).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Dyson has stopped selling its $500 CameraJet, a toothbrush with a camera and AI that aims streams of rinsing fluid. Early buyers reported water getting into the battery compartment, charging problems and brushes that stopped working, one within 24 hours of unboxing. Dyson acknowledges a component issue in some early batches, has told WIRED it has not been recalled, and has declined to say what went wrong or how many are affected (WIRED · Cybernews).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Áttfalt - Orignal Mix — Exos, Octal
Rinse and repeat — the Uncertain Futures desk