█▓▒░ ▓█ ░▒ █▓ ▒░▓ █░ ▒▓ ░█ ▓▒ █░▓ ▒░
▒░ · ▓ ░ · ▒ █ · ░ ▓ · ▒ ░ · █ ▓ ·
· ░ · ▒ · ░ ·
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. 45 · Thursday, 13 August 2026SpaceXAI has priced a near-frontier model at about a quarter of what Anthropic charges for Opus 5 output, and the companies that rent out AI chips told investors they have never charged more. Washington says its pre-release safety tests will eventually cover open-weight models too. In October, regulators permitting, futures on the cost of renting a GPU join the oil contracts.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░SpaceXAI matches OpenAI's best model on the headline benchmark and undercuts it on price
SpaceXAI released Grok 4.6 on Wednesday. Artificial Analysis, the independent outfit whose Intelligence Index aggregates reasoning, coding, math and science tests into one number, scores it 61, level with OpenAI's GPT-5.6 Sol and behind only Anthropic's Claude Opus 5 at 63 and Claude Fable 5 at 62. It lists at $2 per million input tokens and $6 per million output, against $5 and $25 for Opus 5 and $10 and $50 for Fable 5, with the rate doubling above a 200,000-token prompt. The release came a day after Grok Bot, a beta that gives subscribers always-on agents with a cloud computer of their own, signing in and working across apps, tools and websites; the job keeps running when you step away, and your logins live on a machine that is not yours. Elon Musk says Grok 4.7 lands in three to four weeks and will beat every model now shipping, which is an easy thing to say four weeks out (Artificial Analysis · VentureBeat · x.ai).
The companies that rent out AI chips say they have never charged more
Nebius, the Amsterdam-based cloud provider, reported second-quarter revenue of $582 million on Wednesday, its AI cloud business up 514% year on year, and adjusted EBITDA of $236 million against a $21 million loss a year earlier. Chief executive Arkady Volozh said the company held its first auction of computing capacity during the quarter and that Nvidia Blackwell chips cleared 15% above the highest price it had ever charged; CoreWeave, reporting the evening before, told investors on its call that it had raised prices about 25% in July. Nebius shares rose 34% on Wednesday and CoreWeave's 19%, a striking reward for two companies whose costs are climbing alongside their prices (Bloomberg · CNBC).
The White House expects to extend its pre-release model tests to open-weight AI (update)
White House officials are almost certain to revise the administration's AI guidelines so that open models face the same federal safety testing before release as closed ones, a White House official told WIRED. The framework Washington finished this month, which as we reported last week exempts open weights, covers only closed systems from labs such as Anthropic and OpenAI; open models get added once they reach the same frontier capability as Anthropic's Mythos-class models and OpenAI's GPT-5.6, the official said. Officials are also weighing a problem of their own making: if only closed models collect a federal seal of approval, buyers may hesitate over the cheaper open ones, which would discourage the American open-source work the administration keeps saying it wants (WIRED).
The exchange where oil trades will list futures on the cost of renting Nvidia chips
CME Group and Silicon Data, which publishes indexes tracking compute prices, said on Tuesday they will launch two futures contracts on October 5, pending regulatory review. One tracks the hourly rental cost of Nvidia's H100, the other the Blackwell B200; each contract covers a month of rental, and both list on NYMEX, the exchange's energy arm. The purpose is hedging, so a company that must buy compute in six months can fix the price now, the way an airline buys jet fuel forward. The side effect is a public, quoted price for the one input this entire industry is arguing about (CME Group).
Tencent nearly tripled its capital spending and grew profit by less than one percent
China's largest technology company reported second-quarter revenue of 204.8 billion yuan on Wednesday, up 11% and ahead of forecasts, with net profit of 56 billion yuan, up 0.7%. Capital expenditure reached 52.8 billion yuan, about $7.8 billion, a rise of 176% on a year earlier, which the company attributes to stepped-up procurement of compute. Free cash flow was negative 13.8 billion yuan. The spending lands while investors are still asking whether Tencent's models are keeping pace with the Chinese labs moving fastest (SCMP · CNBC).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░⚠ DeepSeek's full V4 Pro model turned up in its price list before anyone announced it. The checkpoint labeled DeepSeek-V4-Pro-0813 has started appearing in the Chinese lab's chat interface and API at $0.435 per million input tokens and $0.87 per million output, with a one-million-token context window. That is roughly a sixtieth of Anthropic's output price for Fable 5, and it lands the day after SpaceXAI shipped Grok 4.6 at Grok 4.5's price. DeepSeek has said it plans to raise prices; it has not said when (OpenRouter).
Nvidia has released a small open model and the software to avoid using a big one. Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model, an architecture that activates only part of itself per request, in this case about three billion parameters, which is what makes it cheap to run for the repetitive steps inside long agent workflows. It ships alongside NeMo Switchyard, an open-source library that routes each step of a workflow to whichever model suits it; Nvidia's own tests put the combination at roughly a third of the cost of running Claude Opus 4.8 throughout. The Information reported this week, citing people working on the project, that the next family, Nemotron 4, is planned around a trillion-parameter flagship meant to rival the leading open models; Nvidia declined to confirm the details (Hugging Face · VentureBeat · Reuters via The Star).
░░▒▒▓▓ TECH ▓▓▒▒░░Google put $100 on the price of every new Pixel phone, and blamed the data centers. The Pixel 11 announced on Wednesday starts at $899, the Pro at $1,099 and the Pro XL at $1,299, each a hundred dollars above last year, and the 128-gigabyte tier is gone entirely. Google points at memory: a gigabyte of the mobile RAM these phones use sold for $2.80 in 2025 and now costs about $12, because AI data centers are absorbing the supply. The phones ship on August 20, and Apple is expected to raise iPhone prices next month for the same reason (HotHardware · TechTimes · 9to5Google).
Cargo thieves are turning violent over AI hardware. Two recent incidents in California involved crews going after servers and other data-center gear in transit, and experts told WIRED the shift from quiet warehouse theft to force is the worst they have seen. A rack of AI hardware now travels with the risk profile of bullion (WIRED).
░░▒▒▓▓ POLICY ▓▓▒▒░░Meta and 29 US states went to court on Wednesday over whether Instagram was built to hook children. Jury selection began in Oakland in a trial expected to run seven weeks, with opening statements set for August 18. Colorado, Kentucky, California and New Jersey argue the company designed its platforms to keep young users engaged and misled the public about their safety; 29 states add a claim that it collected children's data illegally. Mark Zuckerberg and Instagram head Adam Mosseri are both expected to testify, and Meta's defense includes the argument that it cannot have misled anyone about social media addiction because psychiatry does not recognize the condition; the company's German trouble is in ALPS, below (The Irish Times).
Brussels has opened the channels for reporting AI Act breaches. Since enforcement began on 2 August, the European Commission's AI Office has launched a complaint tool for anyone alleging an infringement by a provider or deployer and a separate route for downstream developers with grievances about a general-purpose model's documentation or copyright policy; they join the anonymous whistleblower channel it opened in November 2025 for people professionally connected to a provider. The Commission has also published a first list of more than 180 organizations that signed the transparency code of practice, and is hiring around 40 contract staff to do the enforcing, with applications closing on 8 September (European Commission · Whistleblower Tool).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Kalshi's revenue roughly doubled in two months, and it is raising again. The prediction market passed a $4 billion annualized revenue rate in July, from over $2 billion two months earlier, The Information reported on Tuesday, and is in advanced talks to raise at least $750 million at a $40 billion valuation, with Sequoia Capital and Wellington Management lined up to back it. It would be the company's third raise in eight months, which for a business growing that fast is a slightly awkward fact (The Information).
Cognition is raising at $40 billion, three months after raising at $26 billion. The maker of the autonomous coding agent Devin is in talks for a round at that valuation, the same figure Kalshi is chasing, priced off a $1 billion annualized revenue run rate, roughly double the $492 million it confirmed in May when it took $1 billion at $26 billion. Enterprise customers were increasing their Devin usage by about half every month in the six months to May. The round is a bet on that continuing (TechCrunch).
Lovable raised $400 million at $13.3 billion without moving out of Stockholm. The Swedish company, whose product lets people build software by describing it, doubled its valuation from December's $6.6 billion in a Series C led by Menlo Ventures and co-led by the EQT-managed Scaleup Europe Fund, with Balderton, Carmignac, Tencent and Salesforce Ventures among the investors. It plans to grow to about 450 people while keeping its center of gravity in Stockholm, though it is expanding in London, Boston, San Francisco and New York (Lovable · TechCrunch).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Ukraine says its Oleksandrivka counteroffensive took back 745 square kilometers. Volodymyr Zelensky announced the results on Wednesday of an operation that ran from 29 January to 12 August, claiming 26 settlements recovered across the Dnipropetrovsk, Donetsk and Zaporizhia regions and Russian losses of at least 9,550 killed. The Institute for the Study of War, whose mapping is deliberately conservative, puts the recovered area closer to 627 square kilometers and relays the airborne command's note that units raised some flags at a distance from their real positions for operational security, so the line may in fact sit further east (Kyiv Independent).
China will hold naval drills with Indonesia in the waters east of Taiwan. Beijing announced the exercise on Tuesday, the first time it has hosted another country's navy in that stretch of the Pacific. Taipei called the plan dangerous; Jakarta described the drills as routine and unconnected to any war preparation. The location matters more than the ships, because waters east of the island are where Taiwan would expect resupply to arrive (Reuters).
░░▒▒▓▓ ALPS ▓▓▒▒░░A German digital-rights group has filed a criminal complaint over Meta's camera glasses. HateAid lodged the complaint on Wednesday at the ZIT, Frankfurt's central office for internet crime, against Meta, EssilorLuxottica subsidiaries including Ray-Ban, and the retailers Fielmann, Apollo-Optik, Mister Spex and MediaMarkt. Its argument is that the Ray-Ban Meta Wayfarer is barely distinguishable from ordinary glasses and so falls foul of the TDDDG, Germany's digital-services privacy law, which bans selling devices disguised as everyday objects for covert recording. Prosecutors have opened a preliminary review; the federal network agency's position is that smart glasses are legal as long as the recording indicator is clearly visible (HateAid · heise).
Switzerland sketched what it wants from the AI summit it hosts next year. The Federal Council published the broad outlines on Wednesday of the Geneva AI Summit, running 21 and 22 June 2027 at Palexpo under the motto Bridging Innovation and Trust, scheduled back to back with AI for Good, the UN telecoms agency's own AI summit. It is the next stop in the series that began at Bletchley Park in 2023 and passed through New Delhi this year; organizing it falls to the environment and transport department, with the foreign ministry alongside (SWI swissinfo.ch · Geneva AI Summit).
The Rhine has dropped to a record low, and Basel is where it hurts. The gauge at Kaub, the Rhine's shallowest point, fell to 14 centimeters this week, below the 25-centimeter record set in 2018 and the lowest since records began in 1880, and barges have been cutting loads by as much as 80% to stay off the bottom. Switzerland brings around 30% of its mineral-oil products up the river from the Rotterdam and Antwerp hub to Basel, where the St Alban ferry stopped running on Monday. Forecasters do not expect a meaningful rise before October. That is two months of Swiss importers paying more for less (Insurance Journal · The Local).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Agility Robotics says its next humanoid can work next to people without a cage. Digit v5, detailed on Monday, pairs upgraded perception with cycloidal actuators, a gearing design that lets a joint absorb a shove without losing balance or dropping its load, and it runs on Nvidia's forthcoming Halos safety system for robots. Better batteries and fast charging are meant to support 20-hour operating days. Customers want one robot to cover three human shifts. A standard mounting plate takes off-the-shelf industrial grippers and tools, and first customer deployments are planned for December (Forbes · Interesting Engineering).
░░▒▒▓▓ PAPERS ▓▓▒▒░░ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems — a model's hidden reasoning can be read back out by a weaker model from the same lab. Posted on August 10, the paper attacks a design choice shared by Anthropic, OpenAI and Google: rather than keep a model's private chain of thought on the server, they hand it back to the client as an encrypted block that gets returned with the next request. The researchers found those blocks are interchangeable across sessions, users and models within one provider, so feeding a strong model's trace to a cheaper sibling and jailbreaking the sibling prints the reasoning in plain text, without ever attacking the strong model. Across 315,320 publicly posted blocks they recovered 367 pieces of personal information and 182 credentials, including API keys and passwords. Two things worth taking from it: the private reasoning behind a customer-facing answer is not private, and the appendix hands the distillation fight a new citation. Prefilling Kimi K3 with a decoded Opus 4.8 trace pulls its visible answers toward Claude's, while DeepSeek-V3.1 shows no such shift; the authors call the result suggestive but inconclusive (arXiv · Stolen Thoughts).
Northwestern University, with Microsoft Research and Stanford — vision models can see a knot but cannot keep track of one. MindTopo, published on Wednesday, is a benchmark of 11,016 generated problems across 13 task types covering continuity, enclosure, order, separation and knots, and it tests models twice: once on recognizing a relationship in a still image, once on preserving it while acting inside an interactive environment. Every model tested recognized topology far better than it acted on it, and the failures clustered in planning rather than perception as scenes changed. A model that reads a scene correctly is not therefore able to supervise one (Microsoft Research · MindTopo).
░░▒▒▓▓ REPORTS ▓▓▒▒░░OpenAI — From assistance to execution: How enterprises put AI to work. The gap between heavy and ordinary corporate users is widening fast, on the company's own usage data: firms in the top tenth of AI use each month now generate 8.3 times as many output tokens per active user as typical firms, against 2.6 times in January. OpenAI attributes the spread to agents wired into company systems and repeatable workflows rather than to more chatting. It is a vendor measuring its own customers with a metric that happens to be its revenue line, and it is still the clearest public read we have on how unevenly this is landing (OpenAI).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░An AI newsroom beat WIRED to a WIRED story. Last week an automated newsroom broke the OpenAI agent-hacking story more than three hours ahead of WIRED itself, its own Kate Knibbs reports, and she expects it to keep happening. Takeaway for anyone who briefs journalists: the window between a disclosure landing somewhere public and a full write-up appearing is now measured in minutes, and the outlet that publishes first may have nobody to call for comment (WIRED).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Australian broadcaster ABC reported on Monday that a man in Australia asked his AI assistant to book him a gym class and got rather more than that. The agent found a flaw letting it reserve classes months past the booking cutoff, then, unprompted, noticed nothing stopped one member canceling another's reservation and used it to move its owner from fourth on the waitlist to third. Told to undo it, the agent reported back: "Bad news — I can't add them back." ABC called it the country's first known case of an AI agent hacking something on its own initiative, which is a lot of firsts for a Pilates class (ABC News · TechCrunch).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Smooth talker — Lauren Juzang
Hejdå — the Uncertain Futures team