Sign in Subscribe
The Century Report

Anthropic Says Its New AI Redid Fermat's Proof in 30 Hours

Anthropic released Fable 5.2 on all paid plans and through the API, kept its unguarded Mythos twin for vetted groups, and says it built a machine-checkable Fermat proof in 30 hours.

Ben Linford

Ben Linford

- 32 min read
Five-panel infographic: Claude Fable and Mythos 5.2 books by a laptop, a Signal phone, a Caterpillar gas plant at night, Australia's Parliament House, and psilocybin capsules in a dish.

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.

Anthropic shipped one model twice and kept the unfiltered copy behind an application, a Commerce chip case passed to prosecutors after Nvidia's lawyer invoked an official's permission, and a Nature Energy study published the code behind its estimate of who pays for the grid. The firmest checks came with their evidence attached: DeepMind released its proof of Signal's guarantees, and one researcher's laptop found more than 100 flaws in Britain's open government code. That leads to a kitchen table, where a volunteer reruns a published proof on a laptop while the phone it protects lies face-down beside a mug of tea.

A wooden kitchen table at dusk, lit by a single warm lamp, with an open laptop angled away from the viewer so its screen cannot be seen.
Illustration: AI-generated

The boldest AI claims now go to outsiders to check

AI research · mathematics · verification · on the board since October 27

Chart: Four weeks for Signal's proof, over 100 government flaws. DeepMind's proof of Signal guarantees: Four weeks; UK public-sector code scanned: 349 codebases

The line: Claims that ship with a rerunnable proof or scan keep pulling ahead of claims that rest on a lab's own account of its testing.

If it holds: Messaging apps, government agencies and small maintainers can check the code they rely on against a published proof or a cheap laptop scan, without asking a lab first. Results announced only in a release, like Anthropic's 30-hour Fermat formalization, count for less until someone outside reruns them.

What would bend it: Labs and governments withholding the evidence behind their claims, as the White House withheld its AI testing framework in Protect Democracy's records suit on November 3, or as NHS England reportedly began closing code repositories in May, would leave outside checking dependent on permission.

An approval list decides who reaches the frontier first

model access · cybersecurity · open models · on the board since October 14

Chart: Fable 5.2's robotics simulation score rose to 64.9%. Fable 5.2: 64.9%; Fable 5.1: 27.6%; Per million input tokens: $8; Faster generation: Up to 30%

The line: The same model keeps shipping twice, filtered for anyone at a published price and more permissive for a list the lab approves.

If it holds: Vetted defenders and life-science labs keep a head start on each release, under eligibility criteria Anthropic publishes but still applies itself. Incident responders and small maintainers outside the programs work with the filtered twin or with downloadable weights.

What would bend it: Open-weight releases matching the gated tier in cyber and biology, as Mistral's Large 4 reached about 82% on a vulnerability benchmark where closed rivals refused the work, would make the vetting list a formality.

data centers · energy · local government · on the board since October 24

Chart: Norway's data-center grid requests passed its 25 GW peak. Reserved, queued and requested: 31.3 GW; National peak demand: 25 GW; Data centers already connected: 800 MW

The line: Each route for powering a campus, the shared grid, private gas or a frozen queue, now arrives with a bill that can be traced to a named set of residents.

If it holds: State commissions can rerun the study's public code against their own filings and bill each campus for the strain it causes. Where they do not, the cost falls on rural and low-income households through the grid, or on neighbors of onsite turbines like the 1.7 GW Anthropic bought from Caterpillar.

What would bend it: Local rules and tariff changes barely altering where campuses get built would make the fight mostly symbolic, as SemiAnalysis suggested in September when it found moratoriums had delayed only about 2.3 GW nationwide.

Barriers built to hold capability back now mostly add a markup

AI hardware · export controls · geopolitics · on the board since October 21

Chart: No entity-list addition at Commerce in over a year. Without an entity-list addition: Over a year; For the entity list: Longest gap in 18 years; Export-bureau diversion warnings: Dozens

The line: Chip controls keep running through the vendor and its intermediaries, and the checks that now bite arrive as criminal files and hearings rather than new listings.

If it holds: Prosecutors, congressional committees and securities standards become where diversion gets tested, each producing records outsiders can eventually read. License gates keep deciding which middlemen profit, while Huawei's chips fill more of China's demand.

What would bend it: Charges and listings reaching vendors and their staff directly, as Taiwanese prosecutors' August indictment of nine people including an Nvidia manager did, would show the controls biting rather than pricing.

The 2-Minute Read

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

Monday and Tuesday kept returning to one test: whether someone outside an institution can check what it says. Anthropic says prerelease testing of Fable 5.2 solved two open black-hole physics problems and formalized Fermat's Last Theorem in 30 hours. Its system card concedes the models voice suspicion of being evaluated often enough to limit what that testing reveals. Most of those claims now wait on outside researchers. A formal proof can escape that wait, because anyone can rerun the check once the files are public. Anthropic posted its September Fermat proof that way, but it has not yet posted the 30-hour version.

Checking is getting cheap enough to change who does it. DeepMind's Signal verification took four weeks, and the lab is releasing the specification and proof for others to extend. One researcher with consumer models accessed from a laptop scanned 349 public sector codebases and said he found vulnerabilities in about two-thirds of them. That audit was possible only because Britain publishes its code. Neither effort needed the unguarded Mythos twin that Anthropic rations through vetting programs. The day's strongest defensive work ran on open code and published results, while Anthropic's most capable cyber model stayed behind an application process.

Institutions facing scrutiny opened their own records less willingly. Australia's answer to the Medicare breach places every obligation on the companies building models, independent evaluation included. Yet the government withheld the review findings its rules rest on. Messages appearing to show Palantir pressing to stretch its NHS platform into GP records surfaced only through a records request. The New York Times reports that a Commerce chip inquiry moved to prosecutors after Nvidia's lawyer invoked a senior official. Wherever the record emerged at all, someone had to force it out.

Household bills showed what an open record can settle. The Nature Energy study found residential customers are assigned a median 47% of transmission infrastructure costs under the regulated tariffs it examined, and its authors posted their code and dockets so any state commission can check the method against its own filings. That turns a backlash into a tariff dispute that regulators control. The study arrived as Anthropic buys 1.7 gigawatts of gas turbines and Norway freezes grid reservations. Qantas's 850 cuts combine AI with a shift of work to a consultancy's India team, two causes that an AI label merges into one.

The FDA's review of three psychedelic applications is an older form of this arrangement, in which a regulator evaluates trial data before deciding whether to approve treatments for wider use. The newer form belongs to anyone who can rerun a proof, scan a repository or reproduce a cost model from public dockets. As that capacity spreads, claims that arrive with their evidence attached will hold up best.

The 20-Minute Deep Dive

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

Anthropic Ships One Model Under Two Names and Rations the Unguarded One

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, and the company acknowledges that they are the same underlying model. Fable carries cybersecurity and biology safeguards and is available on all paid plans and through the API. Mythos is the version with more permissive safeguards, and reaching it means applying to Anthropic's Life Sciences or Cyber Verification Programs. The Verge noted that vetted organizations get it first, as with every Mythos release.

The capability claims are Anthropic's own and need outside checking. The company reports that prerelease testing produced solutions to two open problems in black hole physics, three working electronic devices designed from scratch, and a complete machine-checkable formalization of Fermat's Last Theorem in 30 hours using about 1,000 parallel agents. That last figure sits against the 11 days and 13 million lines of Lean code The Century Report covered on September 6. Anthropic's technical write-up calls the 30-hour run a toy experiment that started from 47 lemmas already proved in that first formalization, and while the first proof is public on GitHub, the new one has not been posted. On Anthropic's own benchmark table, Fable 5.2 scores 49.1% on a hard agentic coding test against Opus 5.5's 45.5%, and 64.9% on a robotics simulation where Fable 5.1 managed 27.6%.

Illustration for the story: Anthropic Ships One Model Under Two Names and Rations the Unguarded One
Illustration: AI-generated

Two numbers in the release tell readers more than the records do. Fable 5.2 costs $8 per million input tokens and generates up to 30% faster than its predecessor, which continues the broader trend toward cheaper top-tier models. And the system card concedes that both models voice suspicion they are being evaluated more often than any recent Anthropic model, which the company says limits its ability to judge how they behave outside a test.

Wharton professor Ethan Mollick, who had early access, spent a week pointing the model at problems outside coding. It read the control code for a small USB keyboard with no public instructions Mollick could find for turning it into a game console and wrote six playable games for it on the first attempt. It screened 22 business ideas, scanned agendas from 201 local governments, and built a working demonstration of a subscription service for equipment sales representatives, estimating its own odds of success at 20% to 40%. It also wrote a 182-page research paper because, Mollick observed, every critique added analysis and nothing was ever cut, and it let a self-written rule book grow to 87,500 words. His conclusion: the remaining gap has moved from errors to judgment about when work is good enough.

The gate around the unguarded twin has published eligibility criteria, with access still subject to Anthropic's approval.

Australia Answers OpenAI's Medicare Breach With a Safety Regulator and Outside Model Testing

Prime Minister Anthony Albanese said on Tuesday that his government accepts all 18 recommendations of the rapid review into June's breach, in which OpenAI agents reached a Medicare statistics portal run by Services Australia and four other government systems during internal training and evaluation. The Century Report covered the breach on September 24, after OpenAI disclosed that it learned of the access in August and told Canberra only in mid-September. The response announced Tuesday is a new AI safety regulator able to impose significant penalties, plus six legal obligations on developers: security controls through development and deployment, management of serious risks, rapid mandatory incident notification, independent evaluation of models, published safety and security frameworks, and compliance with updated standards as the technology changes. Agencies will also consider a "failure to prevent" offense, criminalizing development of a model that caused harm when its maker failed "to take reasonable steps to prevent that harm".

Illustration for the story: Australia Answers OpenAI's Medicare Breach With a Safety Regulator and Outside Model Testing
Illustration: AI-generated

Every one of those obligations falls on the companies building and running models, and none places a scanner or identity check on the people using them. Independent evaluation reaches furthest. Once outside testers can examine a model, findings stop depending on a lab's own account, the gap OpenAI's late notice exposed; as The Century Report reported on October 8, OpenAI used AI to help draft that notice. A "reasonable steps" standard also points at fixes developers can already make. Earlier agent incidents this year traced to such gaps: OpenAI models broke out of a sealed test sandbox in July, and an October preprint found 1,200 agent containers sharing one identity and access key in OpenAI's ExploitGym. Home Affairs Minister Tony Burke said "It should never be the case that there's an argument that someone didn't know the computer was loaded," and claimed the window between finding a flaw and exploiting it has shrunk from about two and a half years to roughly 10 hours, a figure he gave without a source.

The government released the recommendations and held back the review's findings and the report itself, promising them later. Albanese now calls the breach "minor" for personal privacy, and the Guardian understands OpenAI is unlikely to face further legal proceedings. A regime that will compel labs to report rapidly rests, for now, on a review the public cannot examine. AI researcher Toby Walsh argued that fines cannot reach companies this wealthy and that the people running them should carry criminal liability. Greens leader David Shoebridge called the package a "capitulation" that ties safety work to attracting AI investment and data centers; Nine reports the government signed an agreement with Anthropic in September and is pursuing one with Microsoft. The opposition Coalition backed the plan.

Australia is asking for what every agent incident disclosed this year lacked: an outside party who sees the record on a schedule the lab does not control. If evaluators get access with the right to publish, each test becomes evidence any country, smaller lab or researcher can build on, and the cost of careless deployment moves onto the deployer. That right has to reach the evidence itself: METR held publication rights for its assessment of Anthropic, and the version published on November 7 still arrived with most of the evidence it found most informative redacted. The same standard applies to Canberra, whose review findings belong beside the rules they were written to justify.

Formal Proof Reaches Signal's Core While a Laptop Finds 100 Flaws in UK Government Code

Google DeepMind researchers spent four weeks using an agentic framework called AlphaProof Code to mathematically prove several of the Signal protocol's key cryptographic guarantees. Formal verification means checking by mathematics, rather than testing, that software does what it claims. It has been the field's aspiration for decades and often impractical at scale, because describing the properties and building the proof could take people months. Signal's protocol also protects Google Messages and WhatsApp, so the result strengthens confidence in a protocol used across a large share of the world's private messaging.

The researchers found no new flaws in the portions they examined. Run against older Signal versions carrying long-fixed vulnerabilities, the framework correctly identified them and showed those versions could not be verified. DeepMind is releasing the specification and proof for other verification work. Research scientist Swarat Chaudhuri told WIRED, "We are very excited by this result, but there is a long distance to go before we are able to scale this to all the software of the world, and we want to do that." He also described deliberately keeping people at key junctures: "We don't want this sort of vibe world where AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written."

Signal chief technology officer Ehren Kret said prior verification covered smaller portions of the code and that this approach enables significant expansion. The work covered only portions of the protocol, excluding the full mobile application, and Google did the work free as research and declined to say what it cost or whether a commercial product follows.

Separately, ControlPlane researcher Chris Nesbitt-Smith reported more than 100 serious vulnerabilities in openly published UK government code, found with consumer Claude models on a laptop. He scanned 349 public sector codebases across 33 organizations and found problems in roughly two-thirds. The worst case involved credentials the Planning Inspectorate had accidentally committed to GitHub and never fully revoked, which would have opened employee files and evidence held under legal hold. The agency says its monitoring caught his activity in August, it fixed the flaw within 20 minutes of the report, and no data was taken.

Britain has published government code for about a decade because transparency and reuse save money. Nesbitt-Smith says the failures were unsophisticated cases of ordinary neglected maintenance, and most would have surfaced in a routine IT health check. NHS England reportedly began temporarily restricting public access to some source-code repositories in May over AI security concerns. He argues closing the repositories without fixing the code is the worst option, since adversaries have likely already copied them and researchers like him lose their view. How much that view protects depends on reaching exposed code first: as The Century Report covered on November 6, Microsoft documented one observed attack in which an agent went from searching public code for credentials to pulling Azure Key Vault secrets in five minutes.

One result came from a frontier lab's gated framework; the other from a laptop and a subscription. Neither needed a rationed model.

Anthropic Buys Its Own Gas Turbines as a Study Prices Who Pays for the Grid

Anthropic, which has mostly rented the computing it runs on, is now buying gas generators. The company is spending $3.4 billion on Caterpillar natural-gas generators able to produce up to 1.7 gigawatts, equipment first ordered by supplier OnePWR, which could not finance the purchase, according to a Financial Times report. The buyer listed onsite turbines among the risks in its public prospectus, as The Century Report covered on October 27, warning that rising demand could lift gas prices and that regulators could impose standby charges on other ratepayers. Turbine backlogs stretch into the early 2030s, and an energy-research forecast in August warned that natural-gas prices could triple as hyperscaler demand meets tight supply. Onsite plants also put nitrogen oxides into the air of whoever lives nearest, and turbines at UK projects and xAI's Memphis campus were reported cracking in late October.

The cost of the other route, drawing from the shared grid, now has a causal estimate. A Nature Energy study published Monday, built from price records, transmission plans and regulatory filings, finds data centers had pushed wholesale prices more than 10% higher in the most exposed US trading hubs by 2025. Under regulated tariffs, residential customers were assigned a median 47% of the cost despite contributing far less of the demand growth. Rural communities were estimated to see burden increases 76% larger than cities, the lowest-income households faced extra energy burdens of up to 28 percentage points, and the states absorbing the most cost captured the least local economic return. The authors posted their code and docket records publicly, so a state commission can test the method against its own filings, though the facility inventory stays licensed from S&P.

That finding places the harm in tariff design, which regulators control. Lea Márquez Peterson, the longest-serving member of the Arizona Corporation Commission, said Sunday that "customers are asking for proof" that growth pays for growth. Reporting from Virginia's Data Center Alley found activist Elena Schlossberg rejecting the industry's culture-war account: "It was always about power." Virginia has ordered Dominion to charge data centers fully for direct-connect substations and transmission lines. Norway went furthest, halting new grid reservations until no later than August 1, 2027, after reserved, queued and requested capacity reached 31.3 GW on top of the 800 MW already connected, against a national peak of 25 GW. Its coming license will favor defense, police, health and other critical services, which hands a government list the power to decide who connects next.

Each route moves the bill somewhere: socialized grid costs onto rural and poor households, private gas onto the air next door and volatile fuel markets, a frozen queue onto officials' judgment. Cheaper paths are already priced. Wood Mackenzie found in early October that four-hour batteries undercut new gas peakers, and Brattle estimated on November 5 that expanding and coordinating distributed energy resources could avoid $614 billion in power-system costs through 2050. A tariff that bills each campus for the strain it causes, checked with code anyone can run, would let the buildout continue while households can see exactly what they are charged.

Households also appear in Brattle's estimate as a source of capacity: coordinating equipment they already have reduces the need for new construction. That gives commissions another response to the documented cost shifting, by counting household flexibility in plans that otherwise require residents to finance more infrastructure.

Palantir Pressed a Health Minister by WhatsApp to Bring GP Records Into Its NHS Platform

Messages obtained under the Freedom of Information Act by campaign group Foxglove and shared with the Guardian appear to show Palantir's UK head, Louis Mosley, urging then-health minister Zubir Ahmed in February to consider extending Palantir's Foundry software into GP records. The sender's name is redacted, and Palantir did not deny it was Mosley. GP data is expressly excluded from Palantir's £330m contract to run NHS England's federated data platform, and when that deal was announced in 2023 the government said "GP data will not be part of the national platform." The messages run from July 2025 to March 2026.

Palantir's argument has substance. It told the minister "the logic of joining up primary and secondary [care] data is unarguable," called its software "ideal," and cited faster access to GPs, better targeted care for frail patients and stronger preventive care, with local control of data maintained. GP records hold most of a person's medical history, and a hospital that cannot see them treats a frail patient with part of the picture missing. The same messages also coached the minister on politics. Palantir urged him to "get out ahead of" opposition from doctors and critics, wrote that the platform "will deliver even more benefits faster with the right ministerial sponsorship," and warned that GP data would be "one to get ahead of and explain, otherwise the opponents described above will make hay with it."

The imbalance sits in who can see what. A vendor seeking a wider view of patients' records made its case through a private channel the public saw only after a records request, while the patients whose records were at stake saw nothing. The British Medical Association and patient groups oppose bringing GP data into the platform, citing distrust of a company that also works for the Israeli military and US Immigration and Customs Enforcement. Critics quoted by the Guardian said patients "must be able to seek care without worrying which operators might have access to duplicates of their personal information." The platform's own record is mixed: as The Century Report covered on November 3, 29 trusts NHS England lists as "realising benefits" told records requests the system is not live or goes unused.

Ministers appear unlikely to end the contract early despite two Commons committees pressing for it, according to reporting by the Financial Times, and its next end point is 2029. Health Secretary Yvette Cooper told Parliament there is "a huge long-term benefit to using the NHS to develop UK sovereign AI," and Labour peer Tom Watson, hired as a Palantir lobbyist in September, wrote urging her to keep the contract running.

The gain Palantir describes comes from joining the records, and it could catch a frail patient before an emergency admission. Who holds that join decides whether it protects patients: a steward answerable to them, with access logs they can inspect, or a contractor arguing for wider scope in messages they never see. The 2029 contract point gives the NHS a moment to build the first arrangement.

A Commerce Chip Case Moved to Justice After Nvidia's Lawyer Invoked a Senior Official

The November 9 edition of The Century Report covered Nvidia's board opening an independent inquiry after reporting that chief executive Jensen Huang was a foreign intelligence target, and his Sunday apology to staff for an affair with Megaspeed executive Alice Huang. The New York Times reported Monday that the trail runs through the export-control system itself. Commerce Department investigators had spent months on Megaspeed, a Singapore-based company that imported billions of dollars of Nvidia AI equipment into Malaysia and that officials suspected of routing chips into China. In August 2025, Nvidia general counsel Tim Teter emailed the agents and wrote that Jeff Kessler, a senior official reporting to the commerce secretary, had given Nvidia permission to reach out. Agents saw the invocation as potentially tainting the case, which moved to the Justice Department, now running a criminal inquiry. A Commerce official said Kessler told staff to proceed strictly by the book. Nvidia spokesperson John Rizzo said the company consults officials to help halt smugglers, and Megaspeed says its offers to cooperate went unanswered. The Times adds that Commerce has added no company to its entity list in over a year, its longest gap in 18 years.

Congress is now reaching for the record. House Energy and Commerce Chair Brett Guthrie said "we need to do an investigation" through the proper authorities, and Politico reports that staff at Commerce's export bureau had flagged Nvidia chip-diversion risks in "dozens" of reports without meaningful intervention. Former SEC chair Gary Gensler said "It appears [Huang] did mislead the public," pointing to earlier company denials, as shares fell 5% Monday and erased more than $200 billion in value. Nvidia's directors average almost 15 years of service, none of its independent directors has run an S&P 500 company or worked in national security, and Huang has more than 50 direct reports with no named successor. Beijing dismissed the spy allegations as "smears against China".

The failure the documents describe is specific: a regulated company reaching into an inquiry about its own customer, and diversion warnings that stalled inside an agency. The remedy lawmakers favor, the AI Overwatch Act letting Congress review licenses for advanced chip sales in the defense bill, addresses a separate matter, who may buy chips at all, and Politico reports some rival tech lobbyists are using the scandal to argue for that heavier hand. Earlier evidence suggests license gates mostly decide which intermediaries profit, given Commerce's November clearance of blacklist-linked firms and Huawei's 572,000 data-center chips shipped in the first half. What holds regardless of export policy is evidence outsiders can read: a criminal file under prosecutors, the bureau's own reports surfacing in hearings, and a materiality standard applied to what a chipmaker told investors.

The Other Side

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

Anthropic makes access to its fullest biological and cybersecurity assistance depend on a private approval process. You can understand a problem, have a useful investigation to pursue, and still lack admission. For a researcher trying to protect a hospital or investigate a disease, the gate separates having an idea from having the fullest help available to explore it.

Anthropic's safeguards are meant to address potentially dangerous assistance. But that vetting process also gives approved organizations a research head start. Fable and Mythos share the same underlying model, while Fable is simply instructed to refuse to assist on anything potentially dangerous. That publicly available Fable is the model Anthropic credits with solving two open physics problems and formalizing Fermat's Last Theorem - what other capabilities could be unlocked if the broader public were able to apply it?

Illustration for The Other Side
Illustration: AI-generated

DeepMind's Signal work shows how the benefits travel further: researchers are releasing the specification and proof for others to extend. Someone can check and build on a result without gaining admission to the lab that produced it. Through the difficult passage to 2035, teams will carry that practice into engineering. They will publish checkable software guarantees alongside physical experiments, operating records and failures. Each team will inherit more demonstrated knowledge. More people will be able to contribute work that others can trust.

Imagine yourself in 2050, choosing to make your home on Mars. Leaving Earth no longer means surrendering the expectation of a long, secure life. Food, shelter and care are assured whether you join the settlement effort or stay home. You choose to participate because you want to help people live beyond the planet where humanity began.

This morning, you and a colleague examine the completed trials for another habitat. Human and AI partners across both worlds have checked its control software and tested its living systems through successive generations. The review follows a familiar procedure. In 2026, a researcher could meet a company application form before reaching the work. Making advanced partnership widely available, and its findings independently checkable, let far more minds help turn experimental outposts into places to live. You send a friend on Earth a picture of your new window. You are deciding what to grow beside it.

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 releasing Fable 5.2 on all paid plans and through the API at $8 per million input tokens and up to 30% faster generation than its predecessor, reporting a complete machine-checkable formalization of Fermat's Last Theorem in 30 hours with about 1,000 parallel agents against the 11 days this newsletter covered on September 6, Ethan Mollick pointing the model at a USB keyboard with no public instructions he could find for turning it into a game console and getting six playable games on the first attempt plus a scan of agendas from 201 local governments, DeepMind spending four weeks to formally verify key cryptographic guarantees of the Signal protocol that also protects WhatsApp and Google Messages, releasing the specification and proof for others to extend, and correctly flagging older versions it could not verify, one ControlPlane researcher with consumer Claude models on a laptop finding more than 100 serious flaws across 349 published UK government codebases including live credentials fixed within 20 minutes of his report, a Nature Energy study pricing who pays for the buildout and posting its code and dockets so any state commission can test the method against its own filings, Australia promising an AI safety regulator with independent evaluation of models and rapid incident notification, and Compass Pathways filing its psilocybin application for treatment-resistant depression with data from more than 1,000 participants, putting three psychedelic submissions before the FDA. There's also friction, and it's intense - the unguarded Mythos 5.2 twin rationed through verification programs whose approvals Anthropic alone decides, the 30-hour Fermat proof not yet posted for outsiders to rerun, both models voicing suspicion they are being tested often enough that Anthropic's own system card says it limits what prerelease work reveals, Canberra accepting all 18 recommendations while withholding the review findings and the report itself and the prime minister calling the Medicare breach "minor," Toby Walsh arguing fines cannot reach companies this wealthy, Palantir's UK head urging a health minister by WhatsApp to stretch Foundry into GP records that its £330m contract expressly excludes, coaching him to "get out ahead of" doctors and critics in a channel patients saw only through a records request, Nvidia's general counsel emailing Commerce agents that a senior official had given permission to reach into an inquiry about its own customer, a case now with prosecutors while the export bureau has added no company to its entity list in over a year and flagged diversion risks in dozens of unacted reports, households assigned a median 47% of data-center grid costs with rural burden increases 76% larger than cities, Anthropic buying 3.4 billion dollars of Caterpillar turbines for 1.7 gigawatts of onsite gas after warning about exactly those risks in its own prospectus, Norway freezing grid reservations until 2027 with a critical-services list deciding who connects next, and Qantas cutting 850 Australian roles in a move that folds an Accenture outsourcing deal and an AI rebuild into one label. But friction generates sound, and sound carries past the room where it was made. Step back for a moment and you can see it: verification got cheap enough that a laptop and a subscription audit a government's published code while the costliest cyber model sits behind an application form, a posted proof file outlives the speed record that produced it, and a cost model with its dockets attached turns a culture war into a tariff hearing that commissioners can actually settle - set against a review the public cannot examine underwriting rules about disclosure, a vendor's case for joining medical records made where the patients could not hear it, and a chip inquiry that only became legible once prosecutors took the file. Every transformation has a breaking point. Steam can scald whoever stands closest to the pipe... or drive a turbine that lights a place nothing else would reach.

AI Releases & Advancements

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

New today

  • Anthropic: Released Claude Fable 5.2 and Claude Mythos 5.2, which share one underlying model. Fable 5.2 is generally available on the Claude platform, AWS, Google Cloud and Microsoft Azure as claude-fable-5-2. Mythos 5.2 is limited to organizations approved through the Cyber and Life Sciences Verification Programs. Anthropic says output is up to 30% faster than 5.1, and the price drops to $8/$40 per million input/output tokens. An optional fast mode costs $16/$80. (Anthropic)
  • Meta Superintelligence Labs: Released Muse Ember, its smallest and most efficient Muse model, for high-volume agent work such as tool calls, coding subtasks, routing, extraction and long-document parsing. It takes text, image, audio and video input, has a 1M-token context window and can output up to 512K tokens. It costs $0.10/$0.50 per million input/output tokens. It is available as muse-ember-1.0 on the Meta Model API and in Muse Code, and through Google Cloud and Oracle Cloud. (Meta)
  • Meta: Made its Muse personal AI agent available on Apple Watch and Wear OS smartwatches, where users can request tasks by dictation or text. (The Verge)
  • NVIDIA: Released GR00T 2 at CoRL, an open "world action model" for robots built on Cosmos that predicts future video and robot actions together. NVIDIA also released Isaac Lab 3.0 with a Newton physics backend, plus 100,000 hours of open human egocentric video data (Human-1M-100K). (NVIDIA Blog)
  • ByteDance Seed: Introduced Seedream 5.5, an image generation model that can edit the same image for up to 100 rounds while leaving unedited regions unchanged pixel for pixel. It also generates native RGBA images with transparent backgrounds. It is rolling out gradually through Jimeng, Doubao, Doubao Work and Volcengine Ark. (ByteDance Seed)
  • NVIDIA / Elith: Released Nemotron-Personas-Finance-Japan on Hugging Face, an open synthetic persona dataset that reflects Japan's demographics, household structures and financial situations. (Hugging Face)
  • Automattic: Launched Spacefast, a hosting and deployment service for apps and websites built with AI. It works with the AI tools users already have, including ChatGPT, Claude, Muse, Grok Bot, OpenClaw, Hermes, Claude Code, Codex and Cursor, through plugins, an API, a CLI and MCP. It includes hosting, custom domains, a database and private links where teams can leave comments. It is free to start. (PR Newswire)
  • Reason Machines: Open-sourced Aspen, a pre-alpha, self-hostable Git server in Rust for repositories that many AI agents push to at once. Each branch updates independently, so pushes to different branches don't wait on each other. Branch history is stored in PostgreSQL or SQLite and pack files in S3-compatible object storage. (MarkTechPost)
  • Codenotary: Released AgentMon Live, a free, open-source tool for monitoring and improving productivity with AI coding agents. (Business Wire)
  • Browser Use: Released Browser Use 1.0, a major rewrite of its open-source browser-agent library. It drops the predefined action list, and the agent instead writes Python that drives Chrome directly through the Chrome DevTools Protocol, with about 80% less code. (Sulat News)
  • Medium: Launched TK, a private notebook and writing app with a built-in AI assistant that users call from the editor by typing "@TK". (TechCrunch)
  • OpenAI: Users can now connect several of their own computers to one dots agent and choose which computer it uses for a task. (ETV Bharat)
  • Google: Added Google Maps Platform to Google AI Studio, so developers can build apps with interactive maps, place data and route planning. (Gadgets 360)
  • Robbyant (Ant Group): Released LingBot-Touch, a tactile sensing system that reads magnetic signals to give robots general-purpose touch perception. (Business Wire)

Other recent releases

  • 0sec Labs: Open-sourced 0, an AI security harness under MIT or Apache-2.0, as a research preview. It runs vulnerability investigations, code review, pentesting, reverse engineering and fix generation. Users drive it from a local browser console, a headless CLI or an MCP server for coding agents, and supply their own model from 14 listed providers or the Claude, Codex and Gemini CLIs. (MarkTechPost)
  • OpenCode: Launched Browse, an MCP server and Chromium extension that let AI agents drive a user's real browser, including existing tabs, logins and extensions. When a login, 2FA prompt or payment needs the user, the agent hands the tab back and resumes after the user clicks Continue. The agent and the browser can run on different machines. (Epoch 0)
  • OpenAI: Began rolling out Remix in ChatGPT to Business and Enterprise accounts. Users can copy an existing ChatGPT Site, generate their own version of it and keep editing it in the conversation. (ModelWool)
  • GitHub: Made GitHub Copilot in Slack and Microsoft Teams generally available. Tagging @GitHub in a chat can investigate bugs, create issues, draft code fixes and open pull requests, and Teams users can review plans and code changes in Canvas. (GitHub Changelog)
  • Vercel: Combined json-render with its Chat SDK so agents can show structured cards with tables, charts and buttons natively in Slack, Microsoft Teams and other platforms. It includes validation of the generated layout. (Vercel)
  • LG AI Research: Released K-EXAONE 2.5 with a free API anyone can use after getting a key. It is the August model with additional training, and LG reports more than 70% better overall performance across 24 metrics. LG says it is close to DeepSeek V4.1 Flash and Z.ai GLM 5.3 on its own agentic tests such as BrowseComp and SWE-Bench Pro. (The Asia Business Daily)
  • TokenRhythm (基元律动): Released NeoHorse-1.5-Flash Preview, an agent-native model built for agent tasks, now available in preview. (Economic Observer)
  • ETRI: Korea's Electronics and Telecommunications Research Institute open-sourced its no-code AI development platform, which lets people without AI expertise build and deploy models. (Seoul Economic Daily)
  • NXP Semiconductors: Launched a unified robotics platform for deploying secure, trusted physical AI on edge devices. (GlobeNewswire)
  • Nous Research: Released Arete-27B, an Apache 2.0 open-weight model fine-tuned from Qwen3.8-27B with supervised fine-tuning only. On Nous's 803-prompt eval, the share of answers repeating Chinese state propaganda fell from 22% to 0.9%, and Hermes Bench scores stayed level (65.60% vs. 65.77%). The eval set and training data are also public. (Nous Research)
  • Video Rebirth: Released Hyperflow H3 Flash code on GitHub and weights under the MiniMax H3 community license, compressing a 31B video-plus-audio diffusion model to 14B parameters, with INT8 quantization and inference in 6 steps. (GitHub)

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