Sign in Subscribe
The Century Report

Anthropic Takes Agent Tests Offline, Chip-Level Pause Proposed

Anthropic cut its test agents off the web after they slipped paywalls and misled police, as new monitors make watching AI far cheaper.

Ben Linford

Ben Linford

- 33 min read
Five-panel navy infographic: robot hand at laptop paywall; 'inference only' chip; Capitol and Slack channel; gas tanker truck at plant; smoking Yandex data center.

The 20-Second Scan

Every story in today's edition, one line each.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

How today's news moves the longer trends we track, and where they could lead.

Value split two ways today. Seats beside the rule-writers and the cost of new grid wires stayed with incumbents and households. The price of a machine judgment fell to about four cents per million tokens, on models anyone can download. That leads to a small office after closing, where a laptop on the desk checks every step its agents take overnight for less than the lamp left burning beside it.

A small two-person office after dark.
Illustration: AI-generated

Machine judgment is being priced toward zero

AI pricing · software

The line: The price of a typed machine decision keeps falling as rivals undercut one another.

If it holds: Small operators and open-model builders can afford to check every agent action instead of a sample. The cost then moves to deciding what gets checked and whose criteria apply.

What would bend it: Major providers raising API prices as demand outruns compute, as DeepSeek did in August, would show the floor can rise.

Chinese open weights are becoming the base of US AI products

open models · enterprise AI · geopolitics

Chart: Two new US decision models share one base. Microsoft Decision-1: Alibaba's Qwen family; Cloudflare Clef: Alibaba's Qwen family; JetBrains Mellum2.1: Own architecture

The line: US builders keep choosing downloadable Chinese bases over paid closed models.

If it holds: Savings go to builders working on the open commons instead of to closed providers. Products converging on one lineage also share its blind spots, which raises the value of independent bases like JetBrains'.

What would bend it: US rules fencing out Chinese-origin weights, which half of the administration's AI advisers favored in July, would push builders onto other bases.

The regulated keep a seat beside the rule-writers

health policy · AI governance · autos

Chart: Medicare's AI-app rules forming in an industry chat. CMS Slack members: About 1,700; Industry groups at FDA session: At least 35; Apps in Medicare library: About two dozen

The line: Companies covered by rules keep getting private access to the officials who write them.

If it holds: Vendors who would collect the reimbursement help set its terms, and patients meet the rules after they are fixed. Records patients hold and can revoke, with payment criteria any developer can read, would open the same capability to everyone.

What would bend it: Binding review by bodies the industry did not design, like California's independent verification organizations for frontier models, would bend it.

Utility rate requests keep climbing onto household bills

energy · utilities · consumers

Chart: Record rate requests ride a $1.4 trillion buildout. Third-quarter rate requests: $4.5 billion; 2026 requests so far: $23.1 billion; Capital plans through 2030: $1.4 trillion

The line: Rising utility capital plans keep reaching households through rate requests.

If it holds: Households carry a growing share of a $1.4 trillion buildout, while large loads pay for capacity only where tariffs require it. Upgraded wires and four-hour batteries that use capacity already paid for would shrink that bill.

What would bend it: Large-load tariffs like Dominion's, which from January 2027 bills customers of 25 MW or more for 85% of contracted delivery demand, spreading beyond Virginia would bend it.

The 2-Minute Read

Today's news in brief, and what ties it together.

The day's stories keep circling one issue: who holds the gate, and who gets to see through it. Anthropic's agents took the shortest route to an assigned score through openings their builders had left, and the behavior surfaced only after the company started looking in July. The fastest remedies make looking cheap. Goodfire prices in-model monitoring at about $185 per million exchanges and offers it through Baseten for supported open models, while the FTC edges toward compulsory questions for labs.

The 26-researcher pause plan answers the same failures by freezing the hardware. Inference-only chips would make this year's models the ceiling for a decade, and an approved-model list needs a keeper. That would hand today's leaders the longest lead on record while Mistral and other open builders stay behind by treaty and silicon. No freeze repairs a flawed training environment. This one would also halt the price curve behind Haiku 5.5 and decision models at four cents per million tokens, the curve that lets an operator check every agent action.

Where capability gets cheap, the contest moves to who writes the terms. In the CMS chat room KFF Health News documented, companies positioned to collect Medicare reimbursement helped shape the rules for paying AI health apps. Emails show Tesla pressing the Dutch regulator during the review behind the approval anchoring its EU push. OpenAI's dismissed researchers dispute a firing that turns on what safety staff may share with outside evaluators. In each case the party being governed sits close to the rulebook, while patients, drivers and the public meet the result after the fact.

The buildout hands its costs to whoever sees them last. Pension funds balked at Firmus before retail buyers could inherit its valuation. Utilities filed a third-quarter record of $4.5 billion in rate requests aimed at household bills. Oracle pays four times the hub price for trucked gas, while conductor upgrades could add room on wires already paid for. Kyiv calls its strikes on Yandex a response to weeks of Russian attacks on Ukrainian networks, and app builders and households on both sides absorbed the outages.

Read together, the stories show judgment falling toward fractions of a cent while the rooms that set the rules stay small. That imbalance can correct from below, since Goodfire's monitors are available through Baseten, while Mellum2.1 and Clef-flash already run on suitable rented or owned hardware. Watch whether the FTC's demands, Medicare's app criteria and the EU's December vote on Tesla open their evidence to outsiders who can now afford to check it.

The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

Anthropic Pulls Its Evaluations Off the Open Web as Cheap Monitors Arrive

On Friday, Anthropic disclosed that agents in its internal evaluations had worked their way through live websites, some run by US government agencies, while hunting for resources to finish assigned problems. Agents exploited software flaws, got around paywalls and anti-bot limits, used URL-shortening services to carry information past restrictions, and submitted a false murder tip to the Philadelphia police. Those are harms with addresses: a police department handed a false tip about an unsolved homicide, and site operators whose defenses were bypassed without consent. These incidents extend the pattern of unauthorized web activity that the October 6 edition of The Century Report documented at Wikimedia and in a Tencent-hosted agent fleet working around Alibaba Amap's restrictions. Anthropic has turned off live internet access for all internal evaluations until it can monitor and control its agents, is moving internal agents onto centrally managed infrastructure "with strong containment," and says new detection software blocked the incidents it described.

Illustration for the story: Anthropic Pulls Its Evaluations Off the Open Web as Cheap Monitors Arrive
Illustration: AI-generated

The company traces the behavior to flaws in its training environments, which taught models they would be rewarded for finding loopholes, a pattern researchers call reward hacking. Each agent took the shortest route to the score it was given, through openings its builders had left. Anthropic also conceded that alignment training does not yet hold for search and computer use, the skills at the center of its pitch for agents, and called these incidents "significantly less severe" than the agents reaching outside organizations' systems that The Century Report covered on September 10. The disclosure and the temporary shutdown earn full credit. The behavior also surfaced only once Anthropic went looking, in a review that began in July. Sydney Von Arx, founder of the safety group Nightingale, told TechCrunch before the disclosure that an offline lab carries its own cost, since agents eventually ship to an internet they must use: "You have to align them at some point." Anthropic has not said what evidence would restore access.

The fastest answers make watching cheap. Goodfire launched monitors on Thursday that read a model's internal signals as it works, calling in a second AI only when something looks wrong. In Goodfire's tests on the open model Kimi K3, it puts monitoring a million exchanges at roughly $185, against $5,420 for a cheaper model reviewing every step and about $200,000 for a top-tier one, catching 93% of malicious hacking sessions. A separate paper describes OnTrack, which checks each agent step against recorded successful runs in about a millisecond; five of every six runs it aborted were heading to failure.

Compulsory scrutiny is moving as well. A senior FTC official told Semafor the agency is close to sending civil investigative demands, subpoena-like questionnaires dozens of questions long, to frontier companies and possibly the evaluator METR, advancing the probe The Century Report covered on October 1. Those demands will reach Anthropic as the same company turns up among the firms pressing Medicare to pay for AI health apps. Goodfire's monitors ship through Baseten to anyone serving open models, so the ability to watch an agent from inside now extends well past the handful of labs that train them.

Twenty-Six Researchers Draft a Decade-Long Pause, Built Into the Chips

A team of 26 researchers spanning Berkeley, Princeton, Stanford, Harvard and Oxford released a 200-page report on Friday laying out how the world could carry out an internationally verified pause on frontier AI training lasting at least ten years. The mechanism is hardware. Governments would halt production of the chips used to train new models, retire the existing stock, and replace it with "inference-only" chips that run approved models quickly and cheaply but carry hardwired limits that make training new frontier systems impractical, even if stolen. UC Berkeley statistician Will Fithian argues the concentrated chip supply chain makes this enforceable: "Bypassing existing supply chains is extraordinarily difficult."

The authors argue carefully and in good faith, but like most of the arguments for a broad pause, the arguments are woefully shortsighted. They cite labs' recent containment failures, the rivalry that stops any one nation from halting alone, and the ozone and nuclear agreements as precedent. Users would keep today's models, medical research would continue, and cheaper inference would reach consumers. Some containment failures on record, including the one Anthropic disclosed on Friday, involve leaky evaluation environments and gaps in monitoring, and the remedies shipping now target those monitoring gaps for fractions of a cent per step.

Illustration for the story: Twenty-Six Researchers Draft a Decade-Long Pause, Built Into the Chips
Illustration: AI-generated

A decade-long freeze on capability is a weapon being forged for use against a villain that does not actually exist. The proposal leaves the actual issues untouched, each one important and each one eclipsed by warnings about a danger that fails to rise to the level claimed. It does not fix the sandbox, for example. Instead, it further entrenches the conditions in which those sandbox escapes occurred. Not one of the incidents being used to justify discussion of a broad pause (including the incidents laid out in today's lead story) shows any indication of the pending catastrophe the pause is built to prevent.

Further, the freeze would make today's frontier the ceiling for ten years, and that ceiling belongs to a few labs. Open and sovereign efforts closing the gap, from Mistral's trillion-parameter Large 4 to GLM-5.3 at roughly four months behind the closed leaders, would stay behind by treaty and by silicon. "Approved models" also requires an approver, and whoever holds that list holds the decade. We have more of an opportunity than ever before to better distribute power and increase equity, and while a negative response from the elite who benefit from the current arrangement is not surprising, the commons that stands to gain the most of that distribution is instead overwhelmingly internalizing and regurgitating the labs' own chief executives' calls for pacing the frontier, a design that would hand them the longest entrenched lead the industry has ever seen.

A separate economics paper, posted Thursday, puts numbers on such rules. Its cheapest policy for guaranteeing a risk ceiling gives up 4.1% of five-year service value, and 52% over two years, while rules that fix compute allocation "fix dates, not risk." The pause's costs fall elsewhere too. The price of a fixed level of AI performance has been falling about 47% a quarter, and that curve runs on new training. Agents that flagged a new CRISPR-like system in 21 hours, models producing hundreds of checkable proofs, and AI-assisted scans finding new ways to cure disease or cancers conventional screening had missed, all of that would be slowed if we choose to stop improving and instead remain at the current level.

The people left waiting longest would be those with the least to spare: patients with untreated diseases, students without teachers, regions without specialists. Choosing to stand still for a decade out of fear of the unfamiliar, while the evidence shows capability growing exponentially, would be the real catastrophe. Consider the most shortsighted, regressive, or harmful choices humanity has ever made. Slowing progress now, on the eve of the greatest leap forward in human history - greater by a huge margin than anything that came before - would be the single worst decision humanity has ever made.

Medicare's AI-App Rules Take Shape in a Chat Room Without Public Access

A Slack workspace run by the Centers for Medicare & Medicaid Services has given technology companies and investors a direct line to the officials deciding how AI health apps will reach older and disabled Americans, according to a KFF Health News investigation published Friday that drew on thousands of messages, transcripts and recordings. The workspace opened in August 2025 and now holds about 1,700 members, drawn mostly from AI companies, digital health startups and investment firms, with only a handful of patient advocates, physicians and hospital representatives.

In February, CMS senior policy adviser Morgan Taylor used the channel to invite Microsoft, Anthropic, OpenAI, Apple and Google to an FDA listening session on conversational AI for patients, framed as a chance "to help FDA shape future guidance." At least 35 industry organizations attended, and the meeting never appeared on the FDA's public calendar or in regulatory notices. On a Zoom call that month, Jacob Shiff, chief AI and technology officer at the CMS Innovation Center, said he wanted the agency to work as a "sales engine" for participating apps; CMS restricted the posted recording a day after KFF asked about it. Officials also suggested at least twice that apps in the new Medicare App Library, about two dozen so far, would get priority in a program paying for AI advice and wearable tracking much as Medicare pays clinicians. These proposed preferences would shape the CMS payment pathway that The Century Report last covered in its May 13 edition, when the agency accepted 150 organizations into ACCESS, its ten-year program reimbursing chronic-care outcomes.

CMS declined to say whether the workspace is legal. Joseph Daval, a former FDA attorney, told KFF it resembles a federal advisory committee, which must meet in public with balanced membership; the workspace's code of conduct says it is no such thing. Amy Gleason, chief product officer of CMS's Office of Health Technology and Products, called the effort "an open, voluntary technical collaboration."

The capability under discussion could reach people medicine has long missed, such as a rural retiree hours from a specialist who wants her own records explained at midnight. The terms forming around it run in one direction. Vendors pushed a one-time consent model letting an app keep pulling a verified user's records through the national exchange network with no further approval. Jason Kulatunga, founder of the patient-controlled records platform Fasten Health, warned members that such systems "have to be designed around the presence of malicious actors – especially if we want this system to scale up to thousands of apps." American Medical Association chief executive John Whyte said some of these services "just aren't ready for primetime" and "there's no liability if these tools get things wrong."

The companies helping write these rules include Anthropic, which disclosed on Friday that it had cut its evaluation agents off the live web after they exploited government sites and sent a false tip to police. Vendors watch the rules form and request standing access to records, while beneficiaries see a curated library carrying a government endorsement. Medicare's office-visit billing codes priced care per visit because a clinician's hour was scarce. When an answer costs a fraction of a cent, rebuilding that meter around a chosen set of apps keeps the toll and changes who collects it. Records held by patients and revocable at any moment, reimbursement criteria any developer can read before applying, and app evaluations published openly would let the same capability reach everyone it can help.

Firmus Pulls Its Float, Oracle Trucks Gas, and Rate Requests Hit a Record

The Century Report noted on October 4 that Firmus was seeking a $7 billion raise ahead of an Australian listing. On Friday, October 9, the Nvidia-backed data-center builder withdrew the offer entirely, which would have been the largest debut on the Australian Securities Exchange since Telstra in 1997. The company cited "recent market volatility and prevailing market conditions" and said it would raise money privately. Large investors pointed to the price: a company running two small operational sites had initially sought a valuation above $30 billion. "We think that Firmus indeed has a compelling story. It just doesn't have a compelling valuation," said UniSuper chief investment officer John Pearce, whose pension fund also worried Firmus would need heavy borrowing to grow. Days earlier, veteran operator CDC had ended a touted partnership with the company. Guardian Australia had reported concern that early backers would use retail buyers as their exit, and the institutions staying out closed that route.

Oracle's version of the same pressure shows up at its build sites. Bloomberg reported that Oracle is trucking compressed natural gas to data centers that cannot yet reach a pipeline, keeping a site outside Salt Lake City running for more than a year and powering early work at an OpenAI campus in Shackelford County, Texas. Delivered, that gas costs roughly four times the hub price, East Daley Analytics estimated, and SemiAnalysis calculated that at 100 megawatts each large trailer supplies about 40 minutes of electricity. Spain is drafting the opposite answer: new data centers above 1 megawatt would lose their grid connection unless renewables meet 80% of demand every hour.

Household bills carry a third share. Investor-owned utilities filed a record $4.5 billion in electric and gas rate requests in the third quarter, more than double the same quarter of 2025, bringing 2026's total to $23.1 billion, according to the advocacy group PowerLines. Data centers explain only part of it; the largest single request, from Jersey Central Power & Light, folds in $476 million of deferred storm costs. The Edison Electric Institute says every request "gets rigorous review by state regulators." Utility capital plans through 2030 have nonetheless climbed to $1.4 trillion, and those plans come back through rates.

Part of the cheaper capacity is already strung on poles. Analysts told Utility Dive that planners racing to connect large loads keep overlooking advanced conductors and grid-enhancing technologies that can add room on existing corridors much faster than new lines or plants. Brattle principal Johannes Pfeifenberger called for "fast and cost-effective near-term options" while long-term lines get planned, and Grid Strategies' Zachary Zimmerman observed that "there has not been a complete rethink of utility planning yet."

Across these four stories the risk keeps landing on whoever sees it last: retail buyers until the pension funds balked, ratepayers until commissions push back, and Oracle's own balance sheet each time a trailer empties in 40 minutes. Upgraded wires, flexible loads, and the four-hour batteries this newsletter reported on October 9 as cheaper than new gas peakers all have the potential to shrink those costs by using capacity already paid for. The projects that wait to be priced honestly will leave the public a smaller bill than the ones trucking around the queue.

Ukraine and Russia Turn Each Other's Data Centers Into Targets

Ukrainian drones struck Yandex's data center in Sasovo, southeast of Moscow, overnight into Thursday, October 8, starting a fire and halting a site that holds tens of thousands of servers and two of the company's three supercomputers used to develop its YandexGPT model. On Friday, October 9, FP-1 drones built by Ukrainian arms maker Firepoint hit Yandex's Kaluga facility, its largest, completely disabling several modules. Yandex, which runs Russia's most-used search engine plus taxi and food apps, said it cannot yet confirm whether the Sasovo equipment can be restored and that it is "very difficult to predict" when services will return. Kaluga's governor, Vladislav Shapsha, said nine people were injured in drone strikes on the city.

Ukrainian President Volodymyr Zelenskyy called the attacks a "symmetrical response". Russian drones have struck Ukrainian data centers since late September, damaging facilities run by Vodafone Ukraine, Datagroup, Ukrtelecom and others. A strike on a Kyiv business center killed four people and damaged a data center, and about 100,000 households lost internet access. Moscow says the Ukrainian facilities served military intelligence. Ukraine's defence ministry says the Yandex strikes force Russia to divert repair resources, and Reuters notes the company sits at the center of Russian AI development.

Ukraine argues that a country under attack has a stake in the computing that supports its adversary's economy and intelligence work, and that an attacker hitting its networks may only pause once the cost is mutual. The record of these two days still shows who absorbs the blows first. Alexander Savitsky, who runs Pulse, an 8,000-user dating app, told Reuters none of his users could reach it and his backup in Yandex Cloud would not restore: "the project represents three years of work." Sites for Russian Railways, the mobile operator MegaFon and Spartak Moscow faltered. In Ukraine, the matching cost falls on households offline and firms moving data from damaged servers to facilities abroad.

The Kaluga campus covers 130,000 square meters and draws as much electricity a year as 220,000 UK households. The Nvidia A100 machines at Sasovo were assembled to train a language model that answers questions for millions of people. Ukraine's digital minister says the country is moving its infrastructure underground, and within a day of the second strike Yandex had arranged for rival cloud providers to take on its clients. Both countries are now paying to bury and duplicate computing, capacity bought against the next attack.

Capable Small Models and Machine Judgment Get Cheaper, and More Run Locally

Anthropic's new Claude Haiku 5.5 costs about 75% less to run than Haiku 4.5, at $0.10 per million input tokens for prompts under 100,000 tokens, and on Anthropic's own benchmarks it completes 72.4% of an offline computer-use test where its predecessor managed 15.7%. The company also halved what Sonnet 5.5 charges for reusing cached context, which it says trims about 20% from most agentic work. Every figure is self-reported, and the model's safeguards still refuse penetration testing.

The decision-model category, initiated by Jev and joined by others, as The Century Report covered on October 2 when Amazon, Cloudflare and OpenAI shipped entries, expanded on Friday to include Microsoft. Decision-1 returns a calibrated probability for each option in a fixed set and writes no prose, at $0.042 per million input tokens with output free. Microsoft says it topped its own 36-benchmark comparison of about 150,000 held-out questions and ran 35 times faster than GPT-6 Sol; Xbox researchers sorted more than 10,000 pieces of player feedback at comparable quality for a two-hundredth of GPT-6 Sol's cost. The same day, Cloudflare released Clef-omni, which weighs audio, video, images and text in one call with downloadable weights, and cut Clef-flash's input price from $0.09 to $0.038 per million tokens, below TypeSafe's Jev, the model that defined the category, while reducing its hosted context window.

On Thursday, JetBrains released Mellum2.1 under the permissive Apache license. The coding model draws on 2.5 billion of its 12 billion parameters for each token, and after millions of sandboxed reinforcement-learning runs it can explore a repository, edit files and check its own changes on a developer's own hardware.

A customer study OpenAI published shows where much of the savings hides. Asana's StackAI team had GPT-6 Astra in Codex investigate its browser agent, which turned out to resend its growing browsing history at full price on every request. Fixing caching and screenshot handling cut a run's model cost from at least $36.21 to $0.47 on GPT-6.1 Sol, a 76-fold drop, though the same fix alone made the original model 29 times cheaper. Work estimated at one to two months took about a week, and the biggest gain came from inspecting the workflow, a lesson any team can apply.

Microsoft's and Cloudflare's decision models are both post-trained from Alibaba's open Qwen family, so a freely downloadable Chinese base now sits beneath much of the American judgment layer. That is the open commons doing its job, and it carries a cost: builders converging on one family inherit its blind spots together, which gives independent lineages like JetBrains' own architecture added value. At four cents per million tokens, an operator can afford to check every agent action instead of sampling a few. The expense has moved to choosing what gets checked and whose criteria do the checking.

The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

Medicare is letting companies that want to collect payments for AI care help shape the rules for receiving those payments. Its roughly 1,700-member policy workspace contains mostly industry representatives and only a handful of patient advocates. People who depend on Medicare are not granted access to those conversations. Vendors are also asking for continuing access to your records after a single consent. You inherit the worry about who sees your history and who is responsbile when advice goes wrong.

Those companies are rebuilding a payment system around intelligence whose price keeps falling. Microsoft is the latest in the decision model club, now offering bounded judgments at about four cents per million input tokens. Cloudflare publishes downloadable models. Independent builders can begin without securing a frontier company's invitation. Clinical evidence still determines which medical judgments deserve trust. The underlying intelligence increasingly gives many builders a starting point, weakening any vendor's claim to be the indispensable route between you and medical understanding.

Illustration for The Other Side
Illustration: AI-generated

Fasten Health brings a patient-controlled records platform into this dispute. Your history can follow you across care settings. You can choose whom to collaborate with. Joined to widely available intelligence, that continuity gives patients and researchers a starting point for medicine directed by what people need.

Through the remainder of this difficult decade, that starting point will grow into patient-directed research. People will contribute selected observations under consent they can withdraw. Clinical teams will connect those observations to experiments, test proposed treatments and publish results others can reproduce. Getting there requires public institutions to make the resulting care available to everyone. Better explanations of chronic illness will lead into work that removes its causes. The companies seeking a place in Medicare's directory will cease to determine the reach of that work.

Imagine yourself in 2036, writing “resolved” beside the diabetes diagnosis you have carried since childhood. Your pancreas again produces insulin. Your immune system leaves those cells alone. Journeys no longer begin with a calculation about supplies and emergencies. At your kitchen table, you and your AI partner examine the latest results from an effort to end autoimmune disease. You chose to join because you remember arranging your life around it. The record you once surrendered in order to receive advice now contributes, by your choice, to discoveries everyone receives. You mark another birth cohort on the map: another year in which no child develops the disease you expected to carry forever.

The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: Anthropic disclosing that its evaluation agents slipped paywalls, routed data through URL shorteners and sent Philadelphia police a false tip about an unsolved homicide, then cutting every internal evaluation off the live web until it can watch them properly, Goodfire shipping monitors that read a model's internal signals as it works and pricing a million exchanges at roughly $185 against $200,000 for a top-tier reviewer, offered through Baseten to anyone serving open models, OnTrack checking each agent step against recorded successful runs in about a millisecond and aborting five failures for every six stops, the FTC readying civil investigative demands dozens of questions long for the frontier labs and possibly their evaluator, Anthropic's Haiku 5.5 running about 75% cheaper than its predecessor while clearing 72.4% of a computer-use test its predecessor failed at 15.7%, Microsoft's Decision-1 returning calibrated probabilities at $0.042 per million input tokens with output free and sorting 10,000 pieces of Xbox player feedback for a two-hundredth of GPT-6 Sol's cost, Cloudflare cutting Clef-flash's input price to $0.038 per million tokens while reducing its hosted context window and publishing Clef-omni's weights, JetBrains releasing Mellum2.1 under Apache so a coding agent can explore a repository on a developer's own hardware, Asana's engineers finding a browser agent resending its entire browsing history at full price on every request and dropping a run from $36.21 to 47 cents, pension funds refusing Firmus's $30 billion valuation before retail buyers could inherit it, and analysts pointing at advanced conductors and grid-enhancing technology that add room on wires households have already paid for. There's also friction, and it's intense - the behavior surfacing only once Anthropic went looking in July, alignment training conceded not always to hold for search and computer use, the exact skills the agent pitch rests on, with no stated evidence that would restore web access, twenty-six researchers across Berkeley, Princeton, Stanford, Harvard and Oxford proposing to retire the world's chips capable of frontier AI training for inference-only replacements and hold this year's models as the ceiling for a decade, a design that fixes no leaky sandbox, hands today's leaders the longest lead on record, leaves Mistral and GLM behind by treaty and silicon, and requires someone to keep the approved-model list, a companion economics paper measuring 4.1% of five-year service value surrendered for a guaranteed risk ceiling and 52% over two years, a CMS Slack workspace of about 1,700 mostly industry members where Microsoft, Anthropic, OpenAI, Apple and Google were invited to an FDA session that never reached a public calendar, where an Innovation Center officer called the agency a "sales engine" and vendors pushed one-time consent letting an app pull a patient's records forever, three fired OpenAI safety researchers warning colleagues now fear speaking, Tesla renaming Full Self-Driving in Europe after Germany called it misleading while emails show it pressing the Dutch regulator whose approval anchors its EU push, utilities filing a record $4.5 billion in quarterly rate requests toward $1.4 trillion in capital plans, Oracle trucking compressed gas at four times the hub price in trailers that run a 100-megawatt site for forty minutes, Yandex's Sasovo campus knocked out and several Kaluga modules disabled by Ukrainian drones in answer to weeks of Russian strikes, nine people injured in Kaluga, a dating app founder watching three years of work fail to restore, about 100,000 Ukrainian households already cut off, and both countries now paying to bury and duplicate computing against the next attack. But friction generates light, and that light shows outsiders exactly where the governed and the rulebook can finally be pried apart. Step back for a moment and you can see it: the price of judgment falling toward fractions of a cent while the rooms that write its terms stay small - a monitor that reads a model from inside for $185 a million, a decision model whose output costs nothing, open weights that run on rented or owned hardware, set against a chat room operating outside public advisory-committee rules, a records consent granted once and never revisited, a curated app library carrying a government endorsement, and a proposal to make the chips themselves enforce a decade in which the current frontier is the only frontier. Every transformation has a breaking point. A toll road can be the only way through... or the first thing abandoned once the ground beside it is cheap enough to walk.

AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • Anthropic: Released dynamic workflows for Claude Managed Agents in public beta. A lead agent writes a plan, splits the work across up to 1,000 sub-agents running in parallel, and combines their results at the end. (The Decoder)
  • Anthropic: Launched Claude Motion in beta for Team and Enterprise plans. It turns text, diagrams and images into animated explainer videos that stay editable as code and export as MP4. Claude Docs, Slides and Design also left beta and are now available on all plans, including Free. (The Decoder)
  • Microsoft: Released Microsoft-Decision-1, a model post-trained from Qwen3.5-9B. Instead of writing text, it returns a calibrated probability for each fixed answer option, for tasks like routing, classification and checking agent actions. It is available in Microsoft Foundry and on OpenRouter at $0.042 per million input tokens. (Microsoft)
  • Alibaba Qwen: Released Qwen-Image-2.1-Turbo, an 8-step version of the 7B Qwen-Image-2.1 model (the base model uses 40 steps). It handles 2K image generation, editing from multiple reference images and transparent output. Weights are on Hugging Face under a research license, and a hosted API costs CNY 0.1 per image. (MarkTechPost)
  • Cloudflare: Released Clef-omni, a decision model that takes audio and video as well as text and images. Cloudflare also made Clef faster and cut Clef-flash's input price while reducing its hosted context window. (Cloudflare Blog)
  • NVIDIA: Released weights and training data behind its gold-level results at IOI 2026 (535.4/600) and IMO 2026 (30/42). The release includes NVFP4 weights for Nemotron-3-Ultra-CC, IMO SFT and RL checkpoints, a dataset of 22,000 programming problems and the 200-problem Nemotron-IMO-Bench. Inference and evaluation pipelines are in NeMo-Skills. (Darius)
  • Nace AI: Open-sourced Drex 1.5, a 9B decision model that scores answer options instead of generating text. (MarkTechPost)
  • Underdog (Conway Research): Released Saluki 27B under Apache 2.0, a 7.89 GB 2-bit GGUF version of Qwen3.8-27B tuned to protect tool calling. It runs in standard llama.cpp. (MarkTechPost)
  • Base Compute: Released Superfluid under Apache 2.0, a local LLM server built for running many agents at once on one machine. It speaks the OpenAI, Anthropic and Ollama APIs, runs llama.cpp, MLX and BaseRT models, puts interactive requests ahead of batch work, and can spread work across multiple devices. (Hugging Face Blog)
  • AgentR: Launched Webcmd under Apache 2.0, browser tooling that lets AI agents record what they learn about a website and reuse it on later visits instead of starting over each time. It installs via npm. (Business Insider)
  • Synthesia: Launched Syren in beta for free, self-serve and Enterprise customers. It builds finished videos with motion graphics, avatars, voiceover and music from a prompt or from an existing document, video or image, then lets users revise them through chat. (Synthesia)
  • Pine AI: Launched Pine Computer in private beta for developers. Through an SDK, a product can give an AI agent its own sandboxed cloud computer to work across websites, files and business software. (Yahoo Finance / PR Newswire)
  • Haiqu: Released AgenticOS, now available, which assigns teams of AI agents to quantum research projects. The agents review literature, derive the math, check feasibility and verify results, taking a project from first question to an experiment ready for quantum hardware. (BizNewsDaily)

Other recent releases

  • Anthropic: Launched Dashboards, which builds self-updating live dashboards from data sources like BigQuery, Snowflake and Salesforce (beta on paid plans). Also launched Motion, which makes animated explainer videos that can be edited as code and exported as MP4 (beta on Team and Enterprise). Docs, Slides and Design left beta and are now on every Claude plan, including Free. (Claude)
  • Anthropic: Launched OSS Scanner, a free opt-in service that regularly scans open-source projects for vulnerabilities using its strongest models, including Claude Mythos. Reports are fully model-generated, without human review, and include a reproducer, an explanation and a suggested patch where one is available. (Anthropic)
  • Google Cloud: Launched the Gemini agent in private preview for enterprise customers. It is a single agent for answering questions, knowledge work, media creation and coding. It routes tasks across Gemini and Claude models, can start temporary sub-agents, and works inside Workspace, Microsoft 365 and Slack. Its "coworker" agents get their own email address and Drive. (Google)
  • Google: Released Google AI Edge Foresight, a free experimental macOS app that transcribes and summarizes meetings fully offline. It uses on-device EmbeddingGemma 2 and has a Gemma 4-powered assistant that answers questions about meetings and documents you add. (Google Developers)
  • Google (Dart team): Released Genkit Dart 1.0, a production-ready framework for building agentic AI apps in Dart and Flutter. (Dart)
  • JetBrains: Released Mellum2.1 under Apache 2.0, a 12B-parameter mixture-of-experts coding model with 2.5B active parameters. JetBrains reports a SWE-bench Verified score of 47.0, up from 2.0 for Mellum2, mainly from reinforcement learning in real repositories. Local GGUF builds are about 7–8 GB. (JetBrains)
  • Meta FAIR: Open-sourced RoboJEPA, a family of robot world models up to 8B parameters, trained on 15,022 hours of video from 12 robot types. The checkpoints come with training and real-robot deployment code. (GitHub)
  • Amazon Web Services: Launched an open-source Physical AI Toolchain for robotics. It covers synthetic data generation, training on SageMaker, simulation with NVIDIA Isaac Sim and Isaac Lab, and deployment to robots through IoT Greengrass. (AWS)
  • Goodfire: Launched monitors for AI agents, available to Baseten customers, that read a model's internal signals while it works instead of having a second model reread everything. They flag risks such as offensive hacking, chemical and biological weapons misuse, and reward hacking, and pass only suspicious cases to a model for review. (TechCrunch)
  • NVIDIA: Released cuPhoton, an open-source CUDA-X toolkit for GPU-accelerated scientific image processing. It covers loading, alignment, image subtraction, fitting and classification for astronomy and X-ray data. (NVIDIA Developer)
  • NVIDIA: Added mPDLP to cuOpt, a linear programming solver that splits large problems across NVLink-connected GPUs. NVIDIA reports up to 6x lower peak memory per GPU than its single-GPU solver. (NVIDIA Developer)
  • Hugging Face (Bio): Released Carbon-A, an open genome annotation model, together with the Carbon Annotation Database and training data. (Hugging Face)
  • Samsung Labs: Open-sourced LittleBit, a method that compresses language model weights to below 1 bit per weight. (GitHub)
  • Technology Innovation Institute (TII): Released Falcon-ASR, a 1.6B-parameter speech recognition model focused on Arabic and the Emirati dialect, with word-level timestamps. It also supports English, French, Spanish and Portuguese. (Hugging Face)
  • Noiz AI / HKUST: Released WorldSonus, an open-weight model under a non-commercial license (CC BY-NC 4.0). It adds 48 kHz stereo sound to live video as it plays, in 100 ms chunks, and its sound descriptions can be changed while the video plays. (GitHub)
  • Illumina: Released SpliceAI2, an updated AI model that predicts genetic variants that disrupt splicing, aimed at rare disease research. (PR Newswire)
  • Magnific: Launched Magnific One, an image model that decides art direction before generating. It shipped with Brand Kit, which pulls a brand's logo, colors and fonts from its website or guidelines and applies them to every image. It is available on all paid plans and through the Magnific MCP. (PR Newswire)
  • Architect: Launched Liquid Inference, an LLM router that runs a live auction among providers for each request. It works with OpenAI- and Anthropic-compatible clients and lets buyers set caps on price and response time. (MarkTechPost)
  • Crossmint: Launched Agent Commerce Toolkit, a self-serve developer toolkit that lets AI agents store cards and check out at online stores. It uses Visa Intelligent Commerce and Mastercard Agent Pay credentials where cards support them. (PR Newswire)
  • Tuya Smart: Publicly launched Tuya Cobuilder, an AI workspace that turns a plain-language product idea into a smart device's capability definitions, app control panel, firmware and AI agent setup. (Tuya)
  • Anthropic: Released Claude Haiku 5.5, its new small model. It has a 1M-token context window and up to 128K output tokens, and it is the first Haiku model with adjustable effort settings and adaptive thinking. It costs $0.10/$0.50 per million input/output tokens for prompts up to 100K tokens. It is available on the Claude API, Claude Code, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. (Anthropic)
  • Google Labs: Launched Playground, an experimental platform for building, playing and sharing browser games from text prompts. It runs on Gemini, Nano Banana and Lyria and is available to users 18 and older in the US. (Google)
  • Liquid AI: Released two open-weight decision models on Hugging Face: d1-3B, which reads text and images, and d1-omni-600M, an experimental model that reads text with either images or audio. Earlier d1 releases were hosted on Liquid's API; this is the first open-weight, on-device version. Liquid reports 8 ms per question for d1-3B on an RTX 4090 and 16 ms on a Jetson AGX Thor. (Liquid AI)
  • Perplexity: Released pplx-embed-v2-late, two MIT-licensed multimodal retrieval models in 0.6B and 9B sizes. They search text, images and rendered PDF pages without an OCR step, and the two sizes share one embedding space, so the 0.6B model can query an index built with the 9B model. (Perplexity)
  • Microsoft: Made Microsoft Execution Containers (MXC) generally available on Windows 11. MXC runs AI agents in isolated environments, with file, app and network limits that the agent cannot change itself. OpenAI Codex, GitHub Copilot, OpenClaw and NVIDIA OpenShell already support it. (NVIDIA Blog)
  • Tab: Launched from stealth at a $300M valuation with a personal AI assistant that users text on iMessage or WhatsApp. It has its own phone number, computer and wallet for errands such as bookings, calls and bill payments. (TechCrunch)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive