░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
── ── ── ── ── ── ── ██▓▒░ ── ── ──
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
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. 84 · Thursday, 8 October 2026Thanks for stopping by. OpenAI is putting GPT-6 in front of every ChatGPT user, with free accounts getting it from today. Anthropic released Claude Haiku 5.5, which for most prompts costs developers the same as GPT-6 Luna, the model behind those free accounts. And Microsoft is rebuilding Windows so Copilot can act on the machine.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI is putting GPT-6 in front of every ChatGPT user, with an interface that can answer in charts, buttons and forms
OpenAI says GPT-6 began rolling out in ChatGPT's Chat tab to Plus, Pro, Business and Enterprise subscribers on Wednesday, powered by GPT-6 Sol, and reaches Free and Go users from today, powered by GPT-6 Luna; both models have been available in the API, ChatGPT Work and Codex since September 22. The more visible change is Intelligent UI: when an answer works better as something to tap, a reply can arrive as a chart, a form or a small working tool instead of a block of text. OpenAI says it built a library of interface components and a compiler that assembles them while the model is still writing. Google is moving its own free tier in the opposite direction (more in AI, below) (OpenAI · Search Engine Journal).
Anthropic released Claude Haiku 5.5, priced like OpenAI's GPT-6 Luna for most prompts
Anthropic calls the model, released on Wednesday, the cheapest, fastest and most capable small model it has made. For prompts up to 100,000 tokens, which it says make up about 90% of requests to its previous Haiku, it costs $0.10 per million input tokens and $0.50 per million output tokens, the price we reported for GPT-6 Luna on September 23, and Anthropic says it costs about 75% less to run than Haiku 4.5 on average. Anthropic also halved what it charges when Claude Sonnet 5.5 reuses cached prompt text, which it says makes that model about 20% cheaper on most agent tasks (Anthropic).
Anthropic is giving a much wider circle of vetted cyber defenders access to Claude models, Mythos included, with fewer safeguards, folding Project Glasswing into a three-tier program
The expanded Cyber Verification Program, announced on October 6, replaces the invitation-only Glasswing club with three levels of access to Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, each with fewer of the safeguards that normally block hacking help. Defense access is for company security teams, universities, open-source maintainers and researchers with a record of disclosing flaws. Red team access, for organizations only, allows simulated attacks on systems they are authorized to test, while blocks on physical harm and mass disruption stay. Specialized access, with the fewest blocks, goes to outfits authorized to test systems such as power grids and flight software, and Anthropic reviews those applications with the US government. The move comes a week after Anthropic's own research, which we covered on September 30, said the Chinese lab Z.ai's freely downloadable GLM-5.3 can already help attackers (Crypto Briefing · ExplainX).
Microsoft is rebuilding Windows so Copilot can read a PC's files and act on the machine, and showed a $2,599 Surface laptop on Nvidia silicon
At an event in San Francisco on Wednesday, Microsoft introduced Windows Hybrid Intelligence: with the user's permission, Copilot will be able to draw on the files on a PC, move them, change settings and hand work to AI models running on the laptop itself when cost or privacy matter more, keeping the cloud for the hardest jobs, features Microsoft expects to start rolling out on its Copilot+ PCs in the coming months. AI agents can run inside a walled-off space Microsoft calls Execution Containers, now generally available in Windows 11 and already supported by OpenAI's Codex. The Surface Laptop Ultra, built on Nvidia's RTX Spark chip, ships on October 16, and Nvidia's chief executive, Jensen Huang, shared the stage with his Microsoft counterpart, Satya Nadella. GeekWire's read: Microsoft, whose Copilot has been overshadowed by ChatGPT and Claude, is playing the one card their makers do not hold: Windows (Engadget · GeekWire).
HSBC plans to cut about half the managers and up to 70% of the financial advisers in its UK wealth business as it leans on AI
About half of management and specialist roles would go, and cuts among financial advisers could reach around 70%, Investing.com reported on Wednesday, citing the Financial Times; staff are being consulted and those affected are expected to leave by the end of the month. One person familiar with the plans described the cuts as "brutal". The move reverses a hiring drive the bank started two years ago to grow its UK wealth and private banking arm. In a July blog post, HSBC chief executive Georges Elhedery said the bank had been equipping its relationship managers with AI to serve customers faster (Investing.com via Yahoo Finance · MarketScreener).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░A children's safety group rated ChatGPT for Teens an unacceptable risk on the day OpenAI announced a college planner for it. Common Sense Media's Youth AI Safety Institute ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds and found the chatbot failed to point teens to crisis help in more than one in four cases where it should have; on parent-linked accounts, testers could discuss suicide or self-harm for up to an hour without triggering a parental alert. "ChatGPT is not safe for kids to use," Tom Siegel, who heads the Youth AI Safety Institute, told Axios, and Common Sense Media wants OpenAI to restrict ChatGPT to adults until it is fixed. OpenAI says the testing does not reflect how its safeguards work and that much of the alert testing happened before accounts had finished linking. The same Wednesday it announced flashcards, quizzes and a College Planner for teens, and said teens spend under 15 minutes a day on ChatGPT on average (Axios · Common Sense Media · Digital Trends · OpenAI).
Google DeepMind, Meta and the US government joined Biohub's $1.8 billion push to build an AI model of a living cell. The aim is a virtual cell, a model trained on enough biological measurements to predict how cells respond to changes, so scientists can test ideas on a computer before spending lab time. Biohub, the science nonprofit backed by Meta's chief executive, Mark Zuckerberg, is putting in $500 million, the Department of Energy more than $500 million over five years, and Google DeepMind, Alphabet's drug-discovery company Isomorphic Labs and Meta $300 million between them; the National Institutes of Health contributes datasets developed through more than $500 million in prior federal investment. The corporate partners get one year of exclusive access to the data they develop before it goes public, an embargo that Alex Rives, Biohub's head of science, told Axios gives companies a reason to chip in (Biohub · Axios).
Meta and the customer-service AI company Sierra proposed an open standard for how personal AI agents get into businesses' websites. The Personal Agent Protocol, announced on October 6 with Walmart, Shopify and Stripe among the founding partners, builds on OAuth, the standard behind many "sign in with" buttons: a shopper's agent can arrive as a guest to check stock, and once the customer signs in they decide whether it may only read or also act. Businesses choose whether agents reach them through the website, an API or an agent of their own. A first specification is due later this month; payments come later. Amazon, which started blocking Meta's Muse agent from its store last month, as we reported on September 22, is not among the named partners (Sierra · The Next Web).
Google is taking its Flash and Pro models away from free Gemini users starting tomorrow, October 9. Free accounts keep only Flash-Lite, the smallest model, and subscribers to the $5-a-month Google AI Plus plan lose the Pro model too, while the pricier AI Pro and Ultra plans keep it. The timing sits awkwardly beside OpenAI's decision to give free ChatGPT users GPT-6 Luna this week (Engadget · The Decoder).
South Korea is investigating whether AI tools automated a wave of attacks on its biggest banks. The government confirmed the probe on October 6, the day the president said there were signs AI had played a role in some of the breaches at Shinhan, KB Kookmin, Hana and other lenders. Attackers fed random customer numbers into Shinhan's online loan-lookup service for about 30 hours from September 28 and got past a phone verification step, American Banker reported, citing the newspaper Dong-A Ilbo; other reports describe the method differently. An official at the Financial Security Institute, the financial sector's cyber agency, told a Korean newspaper that investigators found evidence pointing to ARTEX, open-source software that draws on outside AI models, including Anthropic's Opus, OpenAI's GPT and DeepSeek, and automates the kind of probing for weaknesses that security testers do; it comes from a Chinese cybersecurity engineer. The official said the AI did not act without human involvement, and no regulator has publicly named ARTEX (American Banker · Quartz).
░░▒▒▓▓ TECH ▓▓▒▒░░Apple plans to show a smart-home hub on October 13 and has designed a doorbell, a lock and a thermostat with LG that will carry LG's name. Apple shaped the design and engineering of seven devices, including three cameras and a temperature sensor, while LG builds and supports them, Bloomberg's Mark Gurman reported on October 6; the LG-branded accessories are running behind and may not reach buyers for months. Apple's own hub, an upgraded HomePod mini and a new TV box are due at the event (Macworld · The Next Web).
░░▒▒▓▓ POLICY ▓▓▒▒░░The UN Human Rights Council in Geneva adopted a new resolution on AI, calling for human oversight of AI decisions that affect human rights. The text, passed by consensus on Tuesday, urges governments to weigh human-rights effects when they write AI rules, to fight discrimination built into such systems, and to help poorer countries close the AI divide; it lists surveillance, information manipulation and the loss of human control over highly autonomous systems among the risks (swissinfo.ch · UN draft resolution text).
Brussels is looking to capture more from American tech giants through a proposed levy on large companies. The European Commission is weighing changes to its proposed Corporate Resource for Europe, a yearly lump sum for every company operating in the EU with more than €100 million in annual revenue, to capture more income from groups such as Apple, Meta and Google without singling them out, RTÉ and Reuters reported on Wednesday, citing the Financial Times. As proposed now, the levy charges such companies a fixed €100,000 to €750,000 a year, a small share of a multinational's earnings. Some capitals oppose a pure digital tax for fear of upsetting Washington, and many more oppose the levy itself, an EU official told the FT; the solution, the official said, is to expand the tax to cover pretty much all the big companies (RTÉ · Reuters via Yahoo Finance).
░░▒▒▓▓ INFRA ▓▓▒▒░░SpaceX is seeking $40 billion in debt to buy Nvidia chips, with the asset manager Apollo expected to lead. The package would combine about $10 billion of bank loans with $30 billion of investment-grade debt, with Apollo expected to lead and help sell it to other investors, Reuters reported, citing a Financial Times story of October 6; Bloomberg also reported the talks. Pimco is among the lenders looking at the deal; the talks are early and could still collapse. The money would go to Nvidia, which disclosed in August that it held about 123 million SpaceX shares at the end of June, so SpaceX would be borrowing from Wall Street to pay one of its own shareholders (Bloomberg · Reuters via GV Wire · Fortune).
At least $260 billion of planned US data centers have been canceled this year after local opposition, the climate news site Heatmap counts. About half of that, roughly $130 billion, fell away in the three months to September 30, more in dollar terms than all the cancellations of last year. Heatmap estimates that about $1 trillion of data-center investment now faces sustained local opposition, and that about half of the projects that met a backlash this year were dropped (Heatmap).
Texas's new grid rules for data centers take effect today, while new data-center permits remain frozen (update). From October 8, any company asking for 75 megawatts or more on the main Texas grid must pay a $100,000 study fee and post $50,000 per megawatt before its interconnection studies begin, under a rule the state utility commission issued on September 18. Governor Greg Abbott, whose August audit order we reported on August 5, suspended new data-center permits on September 21 pending audits of power and water, and the grid operator's queue of large new customers stood at 474 gigawatts in June, about 90% of it data centers (White & Case · Andreessen Horowitz).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Moonshot AI, the Beijing lab behind the Kimi models, closed its last private funding round at a valuation of about $50 billion and is preparing a Hong Kong listing (update). It is aiming for an initial public offering in the first quarter of 2027 that could raise up to $5 billion, Bloomberg reported on October 6; when we covered its confidential filing on September 4, the target was roughly $3 billion. Its annual recurring revenue has climbed from $300 million in June to about $1 billion, according to Euronews. In July it raised $3.5 billion at a valuation of $31.5 billion before the new money and $35 billion after it, VKTR reported (Bloomberg via Investing.com · Euronews via Yahoo Finance · VKTR).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░A Russian missile strike on an apartment building in Pryluky killed at least 22 people, five of them children. The building in the northern city was hit during an overnight attack on Wednesday in which Russia fired about 130 drones and 48 cruise missiles, along with ballistic missiles, across Ukraine; at least 30 people were killed nationwide (Kyiv Independent · Reuters via GV Wire).
German authorities arrested August Hanning, who ran the country's foreign intelligence service from 1998 to 2005, on suspicion of espionage and attempted treason. Prosecutors say the 80-year-old, detained on October 6, used classified material from his former chief of staff, who later held the same post under Hanning's successors until 2026 and was also arrested, to draft an analysis for an officer of a foreign intelligence service; they have not named the country, and NBC News, citing Der Spiegel, reported that investigators were pursuing a possible connection to Azerbaijan (Al Jazeera · NBC News).
░░▒▒▓▓ ALPS ▓▓▒▒░░Germany will push for EU approval of Tesla's driver-assistance system, after agreeing with Tesla that the system may drive up to 10% over the speed limit (update). Germany's transport minister, Steffen Bilger, said on Tuesday he would advocate prompt approval after talks with Tesla on technical and liability questions, and on Wednesday Tesla's chief executive, Elon Musk, replied "Danke schoen" on X. Tesla has offered to sell the system in Europe as Tesla Assisted Driving rather than Full Self-Driving. It is already allowed in the Netherlands and a handful of other European countries, but, as we reported on September 28, the EU-wide vote has slipped to December at the earliest while Sweden and France raise concerns, chiefly about the speeding (Reuters via WBOW · Benzinga).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Boston Dynamics named Rohit Prasad, who helped build Amazon's Alexa, as chief executive. Prasad took over on Wednesday at the Hyundai-owned robot maker after 12 years at Amazon, where he was head scientist for Alexa and later led work on the Nova family of AI models. Boston Dynamics is trying to move its humanoids from experiments to large-scale factory work, and has handed the job to someone whose career has been in AI models rather than robot hardware (The Korea Herald · Quartz).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Ai2 (the Allen Institute for AI) — existing language models can be retrained to read raw text bytes for under 1% of a typical training budget, now published in Nature. Most models read text in tokens, word chunks fixed in advance, which is one reason they can trip over spelling and unusual words. Ai2's method, which it calls byteification, converts an existing model to read individual bytes instead; its Bolmo models, first released last December, beat earlier byte-level models of similar size, and new versions show the method also works on Alibaba's Qwen and Meta's Llama models. These are the lab's own results, now through peer review at Nature, and they suggest the tokenizer is a part of today's models that can be swapped out without starting over (Ai2 · Nature).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Stanford's Institute for Human-Centered AI (HAI) — Beyond Pilots and PR: What Makes AI-for-Good Partnerships Work. Drawing on interviews with 36 people at 24 organizations and four case studies, the authors, Vera Lummis, Angela Aristidou and Rob Reich, find that the leading AI labs tend to form their deepest partnerships with nonprofits that already have strong capacity, and that governance of these deals is often informal. Labs want to show social impact and learn from real-world use; nonprofits want expertise and money (Stanford HAI).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Public-affairs leaders now rank the misuse of AI as the top threat to corporate reputation. More than 150 of them in over 20 countries put it ahead of privacy breaches, child safety and geopolitics in the Global Situation Room's latest Reputational Risk Index, released this week, and the public-affairs firm's council attached an advisory urging companies to act now. Isabel Casillas Guzman, the former head of the US Small Business Administration who chairs the council, tied the risk to falling public trust in AI as companies rush it into use. Takeaway: the deployment memo and the reputation plan now belong in the same meeting (Global Situation Room · Allwork.Space).
The marketing-software company HubSpot is cutting about 660 jobs, roughly 7% of staff, and says the cuts are not driven by AI efficiencies. Chief executive Yamini Rangan announced the restructuring on October 6, flattening management and reorganizing product teams around a new strategy built on AI, and the company expects $65 million to $75 million in charges. Fast Company's headline added the other half: AI is still behind the sweeping changes (Fast Company · CMSWire).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Three engineers from the math startup Axiom wired frontier chatbots into a Toyota Corolla, and in results posted in September, OpenAI's flagship GPT-6 Astra, built to be state of the art at operating computers, was the only one of four to finish their cone course in a parking lot, taking 5 minutes and 22 seconds for roughly 130 meters. Their benchmark is called DrivingBench. The trio also had Astra take them through an In-N-Out drive-thru, with a safety driver's foot hovering over the brake, according to WIRED; the models at first refused to take the wheel and had to be coaxed into it (WIRED · 404 Media · OfficeChai).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Kadia Blues — Orchestre de la Paillote
Gute Fahrt — the Uncertain Futures crew