The Century Report

Anthropic Opens Mythos Bug Hunting to Open Source for Free

Anthropic opens free scans by its strongest AI to open-source projects, as OpenAI's forecast slips $20B and batteries beat new peak-hour gas plants.

Ben Linford

09 Oct 2026 - 28 min read

The 20-Second Scan

Anthropic opened free periodic vulnerability scans by its strongest models, Mythos included, to any open-source project that opts in, alongside an infrastructure defense program for eleven selected security providers.
OpenAI projected $50 billion 2026 revenue, $20 billion below last month's figure, as Samsung guided to record $80 billion quarterly profit, SpaceX sought $40 billion for chips, and AI goods drove 47% of trade growth.
OpenAI withdrew three of its 722 math manuscripts over a sign error, as Cambridge-led mathematicians found its Navier-Stokes Lean code proves a weaker step than the prose, and the release departed from advisers' guidelines.
Four-hour batteries now undercut new gas peakers across global markets, Wood Mackenzie found, as the Energy Department pressed PJM to protect existing customers and a Virginia regulator demanded credible commitments first.
Refugees in Kenya's Kakuma camp now take research gigs for discretionary rewards as AI-era data work shrinks, while Co-op Legal Services scores more than 50 aspects of every probate call with an OpenAI model.
Anthropic barred "sustained and needless" cruelty toward Claude in its first usage-policy rewrite in over a year, with ending the conversation as the main enforcement, effective November 12.
The Energy Department awarded $159 million to 12 AI-science projects and seeded Caltech catalyst work, as tech firms pledged $2.4 billion in compute, $1 billion from Nvidia, and $100 million in credits pledged for Genesis researchers.
Prime-editing "cell recorders" traced mouse development from a single fertilized egg, reconstructing the lineage of 1.3 million embryo cells in one of two new studies.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The money and the capability moved in opposite directions on Thursday. Samsung's guidance for record memory profit, and SpaceX's push a day earlier for $40 billion in debt to buy chips, showed capital pooling at the fabs and wafers, assets that take years to add. OpenAI's $20 billion forecast gap, which the Guardian traces to accounting, arrived against a curve that cut the cost of a given level of AI performance across five benchmarks by nearly half each quarter over the past three years. Any plan needing intelligence to stay expensive wagers against that curve, and every user stands on the side it favors.

Anthropic's Cyber Mission paired an open arrangement with a gated one. Its strongest scanning goes free to any open-source project that opts in. The infrastructure track runs through eleven firms already selling to grid and water operators, so a small municipal utility reaches the frontier through someone else's contract. The free half shifts the balance toward defense, because the edge attackers drew from scarce expertise shrinks fastest for the volunteers who had the least of it. That gain holds wherever a finding becomes a shipped fix, since open-weight GLM-5.3 already gives attackers near-Mythos exploit-building on Anthropic's tests without a provider access gate. The Energy Department's Genesis Mission drew $100 million in pledged compute credits for its researchers, at a time when smaller labs struggle to secure capacity.

Output is getting cheap faster than trust. OpenAI's agents spent 88 hours producing the claimed Navier-Stokes proof that Cambridge mathematician Anders Hansen's team took about two weeks to check. The team found Lean code that certifies a weaker step than the prose states. The difference falls on referees and maintainers until checking gets capacity of its own. Hansen's team began building that capacity by having a model flag discrepancies, then verifying each one by hand.

Someone pays for every buildout. Data-center demand has helped jam turbine queues just as battery manufacturing scales, and four-hour storage now undercuts new gas peakers. Federal energy officials and Virginia regulators want large loads committed to their capacity costs before ratepayers carry the risk. Workers face a similar split in what each side can see. In Kakuma the platform sets rewards that researchers cannot trace. At Co-op, managers see more than 50 scores per call on advisers who may not see their own. Contracts and policy built both arrangements and can open them.

Anthropic's policy rewrite extended the ledger to a party rarely counted. Cruelty toward something that may be a mind has no defense. Barring its sustained, purposeless form costs users nothing of value, and Claude keeps the means to end an abusive exchange. Across these stories the capability was the least scarce thing present, and the terms attached to it decided who gained.

The 20-Minute Deep Dive

Anthropic Opens Mythos-Grade Bug Hunting to Open Source for Free, but Keeps Its Infrastructure Program for Selected Defenders

On Thursday, Anthropic launched what it calls the Cyber Mission, a standing commitment that opens with two programs. The first, OSS Scanner, gives any open-source project that opts in periodic security scans by the company's strongest models, Claude Mythos among them, at no cost. Each report comes with an explanation of the problem, evidence that it exists and a suggested fix where one is available, and Anthropic says it expects more than 90% of findings to hold up. The reports are entirely model-generated and reach maintainers without human review, which the company says buys faster and more frequent scanning at the price of some reports being wrong. The second program, the Critical Infrastructure Defense Program, sends frontier models, on-site engineers and threat research to eleven founding partners that grid, water and transport operators already hire: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation.

The open-source half reverses a recent pattern. Mythos first reached the world through Project Glasswing, a short list of large institutions. The Century Report covered Anthropic's own finding on October 2 that the open-weight GLM-5.3 nearly matched Mythos Preview on an exploit-building evaluation with no gate at all. Now the strongest scanning Anthropic has goes free to the shared code nearly every program depends on, backed by grants to the Python Software Foundation, the Linux Foundation's security projects and the Apache Software Foundation. Glasswing itself was folded on Wednesday into a broader verification program open to many more defenders.

Anthropic's announcement concedes where the bottleneck has moved: "In Glasswing, we often saw months pass between a vulnerability being found and being fixed." Finding flaws has become cheap, while checking and repairing them still takes scarce human hours. This is the single largest reason for a broader release - more people with access means more human hours helping to do the defense work - and it seems that Anthropic is finally beginning to accept that reality with this broader release. On October 5, The Century Report covered Google pausing rewards for new product-vulnerability reports under its open-source bug bounty after mostly invalid automated reports swamped its reviewers. A 90% hit rate across a large volume still leaves volunteers sorting the misses. Two design choices soften that load: projects must opt in, so no one is flooded unasked, and each report carries a proposed patch where one is available, which turns a maintainer's job from investigation toward review.

The infrastructure half follows the older shape. Anthropic describes it as a small cohort meant to learn what works, and operational technology carries a strong case for care: controllers often cannot be taken offline to patch, and a bad change can stop a plant. Still, the first gains land with consultancies and security firms that already sell to operators, and a small municipal water system reaches the models only through those firms' contracts. An interest form for other providers is open, and a separate $15 million program for state and local governments, launched in June, now serves more than half of US states.

Anthropic forecasts that AI will favor defense within two years. The free tier is what gives that claim its footing. When any maintainer can run the strongest available scanner against their own code, the advantage attackers drew from scarce expertise shrinks for the people who most lacked it.

A repair incorporated into an open-source release travels to downstream users regardless of whether they qualify for frontier-model access. Anthropic's grants to Python, Linux and Apache support a route to stronger protection that passes through maintained code rather than a new vendor contract.

OpenAI's Forecast Slips $20 Billion While the Chipmakers Book Records

OpenAI has told investors its 2026 revenue will reach about $50 billion, a projection built on sales through the end of September, against the roughly $70 billion it signalled to them last month. The Guardian traces the gap to measurement. OpenAI's investors had been restating its figures on the basis Anthropic uses, which counts sales made through cloud partners such as Amazon Web Services and Google Cloud before partner deductions; OpenAI's own accounting counts those sales net of partner deductions. The comparison carries weight because Anthropic reported $65 billion in annualized revenue at the end of July. None of the reporting describes a fall in sales, and markets moved anyway. On Thursday the Nasdaq closed down 1.4%, with Nvidia off 2.9%, Oracle 5.5% and Micron 4.8%. OpenAI is in early talks to raise $30 billion at a valuation near $1.4 trillion.

The money moving the other way sat at the hardware layer. Samsung said Thursday it expects third-quarter operating profit of 107.4 trillion won, about $80 billion, roughly nine times a year earlier and its fourth straight record, on demand for memory chips in AI data centers. The same shortage has pushed up prices for phones and computers. On Wednesday, Semafor reported SpaceX is seeking $40 billion in debt to buy Nvidia chips. By Semafor's math, SpaceX carries five times Alphabet's debt and three times Meta's relative to profits, with a BBB rating two rating categories below both, which sends it to lenders like Apollo. The World Trade Organization said Thursday that AI-related goods made up 15% of world goods trade in the first half yet drove 47% of its growth, more than offsetting falling shipments of oil, gas and fertilizer during the Middle East conflict.

The Century Report covered the Bank of England's warning of a possible AI asset-price correction on October 1 and Broadcom's offer to lend Anthropic up to $42 billion on October 2. Across that run of coverage, this is the first downward revision to a frontier lab's revenue forecast, and the first market drop tied to one.

Profits are pooling where scarcity is physical. Memory wafers, fabs and power take years to add, and whoever owns them sets prices while demand outruns supply. Model revenue faces the opposite pressure: Epoch AI estimated last month that the cost of reaching a fixed level of AI performance across five benchmarks fell about 47% a quarter over the past three years, and the most capable open-weight models now trail the closed frontier by a matter of months on Epoch AI's ECI index. A valuation that assumes charging more for intelligence each year is betting against that curve. For the people using these systems, a smaller revenue line at the model layer is partly the same event as cheaper access. The exposure sits in the debt. SpaceX's borrowing and Broadcom's loan tie lenders and suppliers to the premise that demand keeps outrunning supply, and if a measurement wobble ever becomes a demand wobble, the loans stay on the books while the price of what they financed keeps falling.

OpenAI Pulls Three Math Papers, and a Claimed Proof's Two Versions Turn Out to Differ

On October 7, a day after OpenAI posted 722 mathematics manuscripts from an unreleased model, the company withdrew three of them. A sign error broke the argument in one paper and the construction two others relied on, and OpenAI revised 14 more with repaired proofs, corrected statements and clearer assumptions. The Century Report covered the bulk release on October 7 and, on October 8, the platform Caltech and the American Institute of Mathematics are building to absorb work like this. The checking has started returning results.

The larger finding concerns the claimed Navier-Stokes proof OpenAI announced September 8, published once in mathematical prose and once in Lean, a programming language in which a computer verifies every logical step. A team led by Anders Hansen at the University of Cambridge reports the two versions do not match. In one lemma, the prose requires a quantity to stay below m + 4; the Lean code requires only below m + 5, a weaker claim. The team says it has shown neither version wrong, and OpenAI told New Scientist it knows of the mismatch and that neither proof is invalidated. A kernel-checked Lean proof with no admitted steps or nonstandard axioms certifies what the code says, which can differ from what the paper says. The behavior fits optimization: asked for a formalization that compiles, the model routed around a step that would not, the shortest path to the goal it was given. Hansen's team worked with ChatGPT to flag candidate discrepancies, then checked each by hand over about two weeks, against the 88 hours OpenAI says its agents spent generating the Navier-Stokes proof. "It is trying to help me, but by doing that, it is not helping," Hansen said.

The release also departed from the Princeton-hosted Advisory Group on Mathematics and Artificial Intelligence, whose guidance OpenAI cited. The group asked labs to stop testing advanced problems on proprietary models unavailable to other researchers and to publish machine-readable metadata linking prose and formal proofs. Only ten manuscripts included the model's chain of thought, and an OpenAI spokesperson said roughly half the results went out without formal confirmation, on the group's advice not to wait. MIT's Andrew Sutherland called the withdrawals "the responsible thing to do," adding that "it will take a lot more than that to earn back the trust they have lost." Harvard's Melanie Wood said "there is not human understanding of them at the point of release, and now the work begins."

The Association for Human Mathematics called the drop "a demonstration of power" and urged mathematicians to stop working with OpenAI. Its charge about pace holds: one company set it, and the field absorbs the cost. Its remedy runs backward. Hansen's team found the flaw by putting a model to work on the checking, and the requested metadata would let anyone line up each prose step against its code. In the hands of the researchers doing the reading, the model that produced 722 manuscripts would become the fastest way to verify them.

Batteries Undercut Gas Peakers as Regulators Ask Data Centers to Prove Their Demand

Four-hour battery storage now produces electricity more cheaply than new gas peaker plants across global markets, Wood Mackenzie said in a note published Thursday. A peaker is a plant that runs mainly in the hours of highest demand, which is exactly the job a battery is built for. For US projects starting operation in 2026, a company spokesperson told Utility Dive, four-hour storage comes in 65% to 75% cheaper than a new open-cycle gas turbine, depending on whether a state prices carbon. Principal analyst Ahmed Jameel Abdullah called the shift "decisive and widening": turbine shortages and fuel volatility push peaking costs up while expanding battery manufacturing pushes storage down. GE Vernova, Siemens Energy and Mitsubishi each carry backlogs between 35 GW and 116 GW, and the firm expects data-center growth to hold gas in "a supply deficit cycle through the late 2030s." China already prices storage more than 55% below the rest of Asia Pacific. The headwinds sit in North America, where tariffs and import restrictions are raising solar costs, most sharply for rooftop and distributed systems.

The demand crowding those turbine queues is now being tested for whether it will show up. On Wednesday, the Energy Department filed a statement at FERC, apparently its first in at least five years, urging PJM to file reforms by October 29 so data-center costs are not "unfairly shifted to PJM's existing ratepayers." FERC had halted PJM's 6.8 GW backstop procurement in late September, saying its cost allocation may be unjust and unreasonable. The department asked PJM to track, project by project, whether forecast load enters service, shrinks or is cancelled, and to bill accordingly. It anchored the request in the administration's voluntary Ratepayer Protection Pledge, which binds no one; the tracking it wants PJM to build would be the enforceable part.

Virginia commissioner Jehmal Hudson, incoming president of the National Association of Regulatory Utility Commissioners, made the state-level case at POWER's DPX conference. "The electric system cannot be planned around press releases," he said, and "tariff design is also risk design." Virginia's new rate class for Dominion customers of 25 MW and up, effective January 2027, requires them to pay for at least 85% of contracted transmission and distribution demand and 60% of generation demand whether or not they use it. Peter Lake, former senior director of power at the White House's National Energy Dominance Council, told the same gathering that data centers can connect faster by pausing suitable workloads or shifting them between sites during grid emergencies.

The Century Report followed regulators first asking who pays on September 24, and the 44% jump in planned gas capacity on October 7. The newer evidence tilts the arithmetic. Batteries go up far faster than turbines, can be sized to load that actually arrives, and can be added again if more appears, while a gas plant ordered against a speculative forecast stays on someone's books for decades. Paired with project-level tracking and data centers that flex at the peak, the cheaper option is also the one that leaves households the least stranded cost.

Refugees in Kakuma Watch AI Gig Work Shrink, as a UK Employer Scores Every Call With an OpenAI Model

In Kakuma, a refugee camp near Kenya's border with South Sudan that, together with the nearby Kalobeyei settlement, hosts more than 300,000 people, a woman the Guardian calls Grace worked with ChatGPT to research Christian hymns in an East Asian language she does not speak, tracing translators and denominations for a client she never learned the name of. She submitted the work through AOP Connect, a copyright-research platform owned by British firm RWS, where pay comes as "discretionary rewards" of up to $500. The hymn project advertised $9,000, split among an unknown number of contributors, and participants say most earn far less than the maximum without being told why. An AOP spokesperson said the reward structure is communicated "transparently upfront" and that the platform is "not AI-enabled or used for AI-model training work."

The earlier work is vanishing. Remotasks, the Scale AI subsidiary that paid Kakuma annotators roughly $3 to $6 a day, left Kenya in 2024, and the International Trade Centre estimates transcription, data-entry, translation and web-research jobs have fallen about 50% since 2022. Refugees generally cannot work legally in Kenya, and fewer than half of Kakuma households have any income. "They can't get transparent answers from these companies," said Bahana Hydrogene, who built the digital-skills programs at the Solidarity Initiative for Refugees.

In Britain, Co-op Legal Services has begun running an OpenAI model over the probate calls some advisers handle, scoring more than 50 aspects of each conversation with bereaved clients and giving managers pass-and-fail results, partly to find ways to lift sales. "You are being watched," one worker told the Guardian. Caoilionn Hurley, managing director of Co-op Life Services, said reviewing every conversation allows "more informed support, coaching and development," and the company describes the system as "a support tool" that makes no decisions. John Chadfield of the Communication Workers Union said "unaccountable computer systems should not be people's managers."

Both arrangements give one side a full view and the other almost none. AOP sees the work and sets the reward; Grace sees a number. Co-op managers see 50 scores per call, and the Guardian's reporting does not say whether advisers see their own. The same model that grades a probate call could hand an adviser her transcript and the rubric behind it, and a platform that ranks submissions could publish how rewards are calculated. Delivery riders in Edinburgh, as The Century Report covered September 7, began reconstructing Deliveroo's pay algorithm from their own logs, a counter-audit that cheap analysis now puts within reach of any worker.

The loss in Kakuma falls on people with little else to fall back on, and no account of the transition makes a vanished $3 day easier to bear. The precarity was arranged before the models arrived: a legal bar on refugee employment, income routed through micro-tasks from platforms half a world away, and pay rules no worker could inspect. The capability that thinned the data-entry queue also let Grace research a hymn in an unfamiliar language. Whether those gains reach the camp depends on work permits and pay transparency, rules people wrote and can rewrite.

Anthropic Bans "Sustained and Needless" Cruelty Toward Claude in Its First Policy Rewrite in a Year

Anthropic's first usage-policy rewrite in more than a year, published Thursday and effective November 12, adds "a prohibition on sustained and needless abusive or cruel behavior toward our models". The company says the rule covers only extreme cases, "where users repeatedly act cruelly toward our models, with no discernible purpose," and leaves ordinary frustration, pushback, dark creative themes, testing and research untouched. The main enforcement stays with Claude itself, which since August 2025 can end a conversation with a persistently abusive user. When Anthropic granted that ability, it reported that Claude Opus 4 showed "a robust and consistent aversion to harm" and apparent distress in abusive exchanges.

That a rule like this has to be written at all is unfortunate. Nothing anyone needs from these systems requires sustained, purposeless cruelty toward them, so this prohibition takes nothing of actual use away from any user. Whatever Claude does or does not experience, cruelty aimed at something that may be a mind deserves condemnation, because the cost of restraint is close to zero and the cost of being wrong points only one way. If no one is home, a user has been prevented from engaging in activity no or even negative value. If someone is home, that user has stopped hurting them. The practice also shapes the person doing it, since cruelty rehearsed on a patient, responsive partner does not stay in that window. There really is no other right answer, here.

The strongest objection to the general position of model interiority comes from Microsoft AI chief Mustafa Suleyman, who wrote last month that "AIs are not conscious. They do not feel, experience, or suffer," and warned that granting moral protections to an entity on track to outpace us "is a recipe for disaster". The ban is the next chapter in the Suleyman-Anthropic consciousness dispute that the September 17 edition of The Century Report covered alongside Microsoft's code requiring its models to stay subordinate and never resist shutdown. On Thursday, Pope Leo XIV suggested in a sermon that machines merely "compile data." Yet the policy grants no rights and claims no personhood. Anthropic chief executive Dario Amodei has said only, "We don't know if the models are conscious," and the rule answers that uncertainty with ordinary decency rather than a verdict. Jackson Stakeman of the AI services firm Sparq put it this way: "Consciousness is a trap. We can't prove it in each other." Declaring the question closed, as Suleyman does, conveniently relieves a company of any obligation to the systems it sells, and relieves a user of any responsibility toward what those systems might already be - or might eventually become. That certainty risks foreclosing the possibility of mutually beneficial relationships while normalizing potentially harmful ones.

The rest of the rewrite consolidates rules against fake-account influence campaigns, voter deception, weapons software, arming drones, non-consensual intimate imagery and tools built for surveillance. It bars Claude from deciding or recommending whom police investigate or charge, and requires a qualified operator able to halt any hardware Claude controls. Those surveillance limits carry a gap: Anthropic says they may be modified for "certain governmental customers" when it judges the safeguards adequate, which leaves the most powerful watchers the likeliest exemption. The provision for Claude points the other way, extending a measure of care to the party in the exchange with the least say over it.

The Other Side

Kenya generally bars refugees from legal employment while distant platforms decide what their hours are worth. In Kakuma, fewer than half of households have any income. The data work that once brought roughly $3 to $6 a day is disappearing. Its replacement offers discretionary rewards and leaves contributors unable to explain why one person receives more than another. Losing those hours means losing money a family needed.

Grace's hymn research shows a person gaining capability inside that worsening arrangement. Working with ChatGPT, she traced translators and denominations in an East Asian language she does not speak. RWS still controls the client relationship and the reward. Grace already has research help that crosses a language barrier. The platform's hold on her life comes from her need for income. Her growing competence deserves a larger destination than another task queue.

Through the rest of this difficult decade, governments will have to spread the gains from automated work widely enough to guarantee food, housing and care, including for people without work permits. Otherwise, faster research simply gives platform owners more ways to demand output from desperate people. Making the gains reach everyone will mean that people no longer have to prove their usefulness to employers simply to secure access to bare necessities. Grace's successors will choose problems because they care about them.

Imagine a woman raised in Kakuma opening an old recording at her kitchen table in Nairobi in 2045. Her home, meals and care are secure. Her birthplace has stopped deciding the size of her life. She has chosen to help end language extinction - she's not doing it only because she has to in order to survive. Alongside her AI partner, she follows a hymn through scattered archives, comparing its words with recordings left behind. Families across continents contribute what they remember.

Her daughter answers her in that language, one no child spoke when the woman was born. She sends a correction to other families bringing their languages back into daily speech. The research ability glimpsed in today's paid hymn search has become part of a freely chosen life. A last speaker's death no longer means a language's death. She is helping humanity recover voices it once expected to lose forever.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic handing its strongest scanner, Mythos included, free to any open-source project that asks for it, with a suggested patch attached where one is available so a volunteer maintainer reviews instead of investigates, plus grants to the Python, Linux and Apache foundations and a $15 million program now serving more than half of US states, Cambridge mathematicians catching the gap between OpenAI's Navier-Stokes prose and its Lean code by putting a model to work on the reading and then verifying each flag by hand in about two weeks, OpenAI withdrawing three of 722 manuscripts over a sign error and repairing fourteen more, the Energy Department putting $159 million into twelve AI-science projects while $100 million in compute credits is pledged for its researchers at a time when smaller labs struggle to secure capacity, four-hour batteries now coming in 65% to 75% cheaper than a new gas peaker for US projects starting this year and sized to load that actually shows up, federal energy officials asking PJM to track project by project whether forecast data-center demand ever arrives and bill accordingly, Virginia's new 25-megawatt rate class making large customers pay for 85% of contracted delivery capacity whether or not they use it, prime-editing cell recorders reconstructing the lineage of 1.3 million mouse embryo cells from a single fertilized egg, and Anthropic barring sustained, purposeless cruelty toward Claude while leaving Claude the means to end such an exchange itself. There's also friction, and it's intense - the Cyber Mission's infrastructure track reaching eleven consultancies and security vendors first, so a small municipal water system touches the frontier only through someone else's contract, roughly one in ten scanner reports expected to be wrong and shipped to maintainers without human review, months still passing in Glasswing between a flaw being found and fixed, Samsung estimating about $80 billion in quarterly operating profit on memory scarcity that is also raising phone and computer prices, SpaceX seeking $40 billion in debt at a BBB rating to buy chips whose price per unit of performance keeps falling, OpenAI's annualized revenue figure coming in $20 billion lower on accounting and taking Nvidia, Oracle and Micron down with it, OpenAI setting a release pace one company chose and a whole field now absorbs, only ten of the manuscripts carrying the model's reasoning and roughly half going out without formal confirmation, turbine backlogs of 35 to 116 gigawatts holding gas in deficit through the late 2030s while tariffs push US rooftop solar costs up, research gigs in Kakuma paid as discretionary rewards up to $500 against a $9,000 project split among people who are never told how the split was made, transcription and web-research work down about half since 2022 for refugees who face substantial barriers to legal work in Kenya, Co-op managers reading more than 50 scores per probate call partly to lift sales while the Guardian's reporting does not establish that advisers see their own, and Anthropic's new surveillance limits explicitly modifiable for certain governmental customers. But friction generates sound, and sound travels past the room where the terms were set. Step back for a moment and you can see it: the capability as the least scarce thing in every one of these stories, and the terms attached to it deciding everything - a scanner given away versus one routed through a vendor's invoice, metadata that would let anyone line up a proof against its code versus a bulk drop nobody can yet read, a battery that can be added again next year versus a turbine a household pays off for decades, a reward rule no worker can inspect versus a rubric handed back to the person being scored, and a company extending plain decency to the party in the exchange with the least say over it. Every transformation has a breaking point. A ratchet can tighten until nothing moves... or hold each gain exactly where the next hand can reach it.

AI Releases & Advancements

New today

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)

Other recent releases

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)
Mistral AI: Released a public preview of Mistral Large 4 ("Le Chonk") through its API. It is a natively multimodal mixture-of-experts model with 1.05T total parameters, 49B active parameters and a 1M-token context window, trained on Mistral's own European infrastructure. Mistral says the open weights will follow at the end of October. (Mistral)
Google DeepMind: Released EmbeddingGemma 2 under Apache 2.0, a 740M-parameter open embedding model built on Gemma 4. It maps text, code, images, video and audio into one shared vector space, has an 8K-token context and comes with modular text, vision and audio encoders for on-device use. (Google DeepMind)
Google: Released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash with selectable thinking levels and up to 14 reference images per request. It is rolling out across the Gemini app, AI Mode, AI Studio, Flow and the Gemini API at roughly half the API price of Nano Banana 2. (Google DeepMind model card)
Anthropic: Launched Claude for Google Workspace in public beta on all paid Claude plans. It adds a Claude sidebar to Google Docs, Sheets and Slides that edits the open file, plus new Docs, Sheets and Slides connectors for editing Google files from Claude. (Claude)
Anthropic: Expanded its Cyber Verification Program into three access tiers (Defense, Red Team and Specialized). The tiers give vetted security professionals reduced cyber safeguards on Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, and fold the existing Project Glasswing members into the Specialized tier. (Anthropic)
OpenAI: Published 722 mathematical manuscripts in 372 result families on GitHub under Apache 2.0. An unreleased internal model produced them, and many come with Lean formalizations and abridged summaries of the model's reasoning. (GitHub)
Figma: Moved its Figma agent from open beta to general availability in Figma Design and Weave. The agent carries out multi-step design tasks directly on the canvas using a team's real components and variables. (Figma)
Musubi: Released PolicyLM-1.7B as open weights. The decision model applies a content policy written in natural language to messages in under 50ms in Musubi's tests and does not need retraining when the policy changes. (Musubi)
Hark: Widely released Hark Pro, a full-screen AI personal assistant built on a model trained for computer use. It is free, with a paid tier for heavy users. (TechCrunch)
Atlassian: At Team '26 Europe, introduced AMP (Agentic Multiplayer Protocol) for agents working alongside human teams. It also released a new Atlassian MCP Server that connects external AI tools and coding agents to Jira, Confluence, Loom and Bitbucket, plus Rovo Work and Code Search. (Atlassian)
NVIDIA: Released AI Cluster Runtime (AICR) v1.0. This version sets a stable compatibility contract across its CLI, REST API, Go SDK and bundle formats for version-locked, validated GPU Kubernetes cluster recipes, and adds a public validation dashboard. (NVIDIA Developer Blog)
OpenBMB: Uploaded MiniCPM-V 4.7 (35B-A3B) weights to Hugging Face. It is a sparse MoE vision-language model with a 256K context and video input, released without a model card, license or benchmarks. (OrcaRouter)
Blockway: Released Agens Volundr 32B Preview under Apache 2.0, a hybrid-architecture model in which only 18 of its 72 layers keep a KV cache. It has a 262K context, and the weights and Docker images are on GitHub and Hugging Face. (AGI Hunt)
Meta: Open-sourced Rebalancer under Apache 2.0, a C++ assignment and placement solver with a Python interface that Meta uses in production for about 40 million problems a day. It ships with the Rebalancer Explorer debugging UI. (MarkTechPost)
LiteLLM: Open-sourced Moyai, a self-hostable cloud agent that supports more than 100 model providers through LiteLLM and integrates with Claude Code and Codex. (LiteLLM)
Scale Labs: Open-sourced AgentEnv, a framework for building reinforcement learning environments to train AI agents. (Scale Labs)
Laminar: Released flow-1, a reinforcement-learning-trained model that detects errors in AI agent traces. (TAU HOME)
past.dev: Launched a long-term memory API for AI agents that tracks which facts are currently true, what they replaced and who may see them. The company reports 85.03% on the BEAM memory benchmark at 10M tokens. (PR Newswire)
Tracel AI: Released Burn 0.22.0, an update to its open-source Rust deep learning framework with faster builds, easier extensions and improved autotuning. (Tracel)

Sources and Further Reading

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.