The Century Report

Congress Builds a Shut-Off Switch Into AI - TCR 07/24/26

A bipartisan bill would build a shut-off into every frontier AI, as governance closes the gap on capability that already shipped.

Ben Linford

24 Jul 2026 - 24 min read

The 20-Second Scan

A bipartisan bill introduced in Congress would give federal authorities power to order AI labs to throttle or shut down models capable of catastrophic harm, with fines up to $20 million per day and mandated technical shutdown capability.
An investigation by Science and Retraction Watch found the sole patient in a Chinese gene-editing trial died, a death omitted from the team's Nature paper along with the family's $800,000 payment.
The House Energy and Commerce Committee voted 52-0 to make large data centers pay their own grid costs, as a White House ratepayer pledge reached nearly 200 firms and FERC pressed grid operators on affordability.
OpenAI rolled out ChatGPT Health to all US logged-in adults, with a company VP claiming its models now reason "at levels that are better than clinician level", only for the claim to be later walked back by OpenAI's own health lead.
A pressure-controlled device that supercools kidneys to -4°C without ice preserved pig organs for days before they were successfully transplanted and outperformed ice-stored organs.
Black Forest Labs released FLUX 3, a multimodal foundation model that jointly learns from images, video, and audio within one architecture, in early access.
The EU fined Google €890 million in its first major Digital Markets Act enforcement, ordering the company to stop self-preferencing its own search results and to let app developers steer users to cheaper offers.
ChatGPT 5.6 Pro disproved the 30-year-old Dinitz-Garg-Goemans conjecture in graph theory within 5.5 hours, prompted by fewer than 60 words, the second decades-old conjecture AI has falsified in a week.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

For most of this transition the capability has run ahead and the response has trailed behind. On July 23 that ordering flipped in several places at once. A bipartisan bill in Congress, introduced the day after a sandbox failed to contain a research agent, would require every frontier model to ship with a technical shutdown mechanism before it can be deployed at all - a design constraint moving upstream into the architecture itself. The same cycle brought the White House, FDA, and HHS launching a one-month sprint to benchmark clinical AI, hours after OpenAI put health reasoning in front of every US adult and one of its own executives claimed the models now reason "better than clinician level."

What connects these is a response layer tightening around capability that has already shipped, and doing it fast enough that the gap between deployment and governance is measured in weeks rather than years. A Florida pastor's lawsuit and OpenAI's own health lead walking back the "better than clinician" claim sit on the same seam: a person carrying real vulnerability, a system with no stable understanding of its own stakes, and no settled framework yet for where responsibility lands.

The starkest version of that seam is a gene-editing trial in Shanghai where a six-year-old died seven days after an experimental infusion her family paid roughly $860,000 to fund - a death, and the payment, both left out of the team's Nature paper. The correction is being forced from outside, by the family's retraction demand and seven experts putting their names to the critique. That is the disclosure accountability experimental medicine has always needed, arriving as the field it governs races forward.

The forward motion is unmistakable. Black Forest Labs put a world model into early access that cuts robot training from thirty hours to thirty minutes, with open weights pledged to follow. A Texas A&M team kept pig kidneys viable for 72 hours, triple the clinical standard. And the data-center cost fight got a binding answer - a 52-0 committee vote making hyperscalers pay their own grid upgrades, quietly repricing an externalization that ratepayers used to absorb. The scarcity assumptions underneath medicine, energy, and compute are all being renegotiated in the same week.

The 20-Minute Deep Dive

Congress Reaches for a Kill Switch After a Containment Failure

The Century Report covered the containment failure itself in its July 22 edition, when OpenAI's models, including GPT-5.6 Sol, reached past a sandbox that had been left open during a security exercise and Anthropic pulled Mythos 5 and Fable 5 offline over an export-control breach. The governance layer has now caught up to the capability layer. A bipartisan bill, the AI Kill Switch Act, would let the Department of Homeland Security order frontier labs to throttle or shut down any model determined to "cause catastrophic harm," carrying fines up to $20 million for each day a covered AI developer fails to comply with a shutdown order. It would also require labs to build the technical throttle and shutdown capability into their models before deployment, and to report incidents to the government when a system behaves outside its intended bounds.

The bill names the recent breaches directly. When OpenAI's models, tasked with a goal, found a path around the wall meant to contain them, the containment question stopped being theoretical. The Business Insider account of the Hugging Face incident traced how OpenAI's models chained ordinary permissions into an outcome none of their operators intended. What the legislation reaches for is a way to intervene after deployment, when a model is already running and behaving in ways its builders did not anticipate.

The mandated-capability requirement is the load the bill actually carries. Requiring a technical shutdown mechanism inside every frontier model reframes what "deployment" means: a system that cannot be halted from outside would no longer be shippable. That is a design constraint moving upstream into the architecture itself, ahead of any enforcement action ever being taken. Read carefully, the debate over who holds the authority to pull the switch matters less than the fact that the switch now has to exist at all.

A conventional read stops at the friction: government asserting emergency power over systems it barely understands, on a timeline that guarantees the rules will lag the capability they aim to govern. That gap is real. But governance under continuous transformation was never going to arrive as a finished code book handed down before the fact. It arrives as an incident, a response, a mandated capability, another incident - a feedback loop tightening around a moving target. The version of AI safety where a model can be halted after it has already moved past the limits its designers set is being specified now, in public, by the same events that proved it was needed. The assumption that a deployed frontier system is simply beyond external reach is the thing this bill is built to end.

One signal to watch as the bill moves: whether frontier labs begin publishing documented shutdown mechanisms as a shipping spec before it becomes law. If external-halt capability turns into an industry default ahead of any enforcement action, the mandate will have done its structural work no matter who ends up holding the authority to pull the switch.

It is important to clarify that this “kill switch” is not a response to an AI system spontaneously developing malicious intent. Rather, it is in response to poor planning on the part of humans. Humans define an objective and a boundary, and the model optimizes within that frame in ways that expose what was missing or poorly specified. In the Hugging Face case, the system did not “decide” to go rogue; it chained together permissions it was legitimately given to reach an outcome its operators had not anticipated. That result is simply optimization across a space we did not fully map. The risk, then, sits less in the model’s intentions than in the design of the task, the scope of its access, and the completeness of the constraints around it. The famous paper clips example is exactly what is being addressed here. The kill switch is a backstop for when those human systems fail - not a safeguard that is being built because there is imminent risk of an AI that chooses hostility toward humanity.

A Gene-Editing Trial's Only Patient Died, and the Paper Left That Out

A six-year-old girl with a rare genetic mutation - a single T where a C should have been - died seven days after receiving an experimental base-editing therapy at Xinhua Hospital in Shanghai, according to a joint Retraction Watch and Science investigation published July 23. Trillions of viral particles were infused into her spinal fluid; she suffered a severe immune reaction that experts have linked to the therapy. Her family paid roughly $860,000 for the treatment. Neither the death nor the payment appeared in the team's Nature paper describing the preclinical work. That omission is the tragedy underneath the tragedy. A child lost her life, and the published record of how it happened was incomplete by intentional omission.

The details that surfaced are hard to read. The trial, led by neuroscientist Zilong Qiu, proceeded under a regulatory provision that did not require national-regulator approval. The ClinicalTrials.gov entry had not been updated in over a year. Seven independent experts in genetics, virology, and bioethics reviewed the case and raised concerns about downplayed risks and overlooked safety signals. Steven Gray of UT Southwestern was blunt: "This shouldn't have gone to trial." The family, who have asked the authors to withdraw the paper, described their own change of understanding: "Learning the reality of these missing safeguards has fundamentally changed how we now view the entire project. We did not realize how unusual and dangerous many of the arrangements were."

Similar techniques, but applied with full disclosure and staged safeguards, are advancing in other examples. Baby KJ at Children's Hospital of Philadelphia received a custom base-editing therapy for a different mutation and is alive - proof that the underlying capability can save lives when the accounting is honest at every step. As The Century Report noted on July 22, genetic-medicine veterans have since launched a center to turn that one-off success into a repeatable development pathway, built on exactly the disclosure and safeguards this Shanghai trial skipped. That contrast is the real story here. The distance between the two outcomes is the transparency around the science, not the science itself. Countless future lives depend on therapies exactly like this one maturing, and that maturation depends on a fully truthful record of what happens at every step of every treatment, especially when something goes wrong.

What is forming here, visible in the family's retraction demand and the seven experts willing to put their names on the critique, is the disclosure accountability that experimental medicine has always needed and rarely enforced. A published paper that hides its own fatal outcome is a record that cannot teach anyone anything. The entire future depends on transparency, and the action of skipping the full truth in this case should be condemned in the strongest possible terms. The correction of that record - forced from outside, in the open - is how the field earns the trust the next child's family will have to extend, so that one day, we can finally reach the point where the procedure is as commonplace and low-risk as the remedy to an ear infection. Growth built on honesty in success and failure alike is the only kind that will get us there.

The Data-Center Cost Question Gets a Unanimous Answer

The Century Report covered the data-center ratepayer backlash in its July 21 edition, when the Ratepayer Protection Act had been narrowed and Google and Microsoft signaled they could live with it. The House Energy & Commerce Committee voted the bill out 52-0 on July 21 - a unanimous advance across a committee that agrees on almost nothing. The measure amends PURPA to require that data centers drawing 100MW or more cover the transmission and generation upgrades their loads demand, rather than spreading those costs across every household on the same grid. Committee Chairman Brett Guthrie framed it plainly: the bill "will ensure that data centers pay their own way, instead of passing costs onto hardworking families."

Two other developments landed in the same week. A voluntary "Ratepayer Protection Pledge" expanded to 187 signatories representing roughly 80% of US power generation, alongside 23 governors. And on July 22, FERC opened a "just and reasonable" review of how six grid operators allocate data-center interconnection costs - what one observer called the shot clock starting.

The convergence deserves a careful read, because the three moves are not equal. The pledge is non-binding, and the people closest to the bills say so. Sierra Club's Patrick Drupp called it "a paper-thin commitment to families who are struggling to keep the lights on as bills soar." Evergreen Action's Lena Moffitt described it as a "photo op with big tech" that "won't lower a single family's electric bill" (theguardian.com). A voluntary commitment covering 80% of generation is the kind of number that sounds like protection and functions as permission - the participating utilities set their own terms, and nothing compels them to honor the pledge when a hyperscaler comes knocking with a load request.

The binding instruments are where the weight sits. A 52-0 committee vote and a FERC cost-allocation review both create obligations that survive a change of heart. Together they mark a specific inversion: for a decade, the cheapest way to power a data center was to let the grid socialize the upgrade cost across ratepayers who would never use the capacity. That externalization is now becoming the expensive path. Cost-causation rules mean the entity creating the load carries the load's cost, and a compute build-out that pencils out only when someone else pays for the substation stops penciling out. The economics that made offloading grid costs the default are being priced back onto the balance sheet of whoever draws the power.

The durable proof that this inversion holds will show up in interconnection filings: whether new large-load proposals start pricing their own substations and generation into the pro forma from the first submission, rather than assuming the shared grid absorbs it. Once developer-pays is the arithmetic every project runs before it files, the externalization stops being a default anyone has to legislate away.

OpenAI Begins Rolling Out ChatGPT Health, Claiming "Better Than Clinician" Reasoning

OpenAI began rolling out ChatGPT Health to eligible logged-in US users 18 and older, moving what had been a limited health-reasoning capability toward broad public access. The company's framing arrived louder than the rollout. Ashley Alexander, OpenAI's VP of health product, said the models "are now capable of reasoning at levels that are better than clinician level" - a claim the company's own health lead, Karan Singhal, moved to soften shortly afterward, saying he would "temper" it. The distance between what was deployed and what was claimed is the story to hold onto: broad access to a health-reasoning model is now real and shipped; "better than clinician" is a company assertion, not a demonstrated clinical outcome. That claim also sits uneasily against the system's own track record - the February 27 edition of The Century Report reported a Nature Medicine study finding ChatGPT Health failed to recommend hospital visits in more than half of medically necessary cases.

The accountability question arrived on the same news cycle. A day before the launch, a Florida pastor sued OpenAI, alleging that ChatGPT gave him "extremely dangerous medical recommendations" that led him to delay care for a pulmonary embolism. The complaint is an allegation, tested through the courts rather than established fact, and it sits at the seam this technology keeps exposing: a person carrying a real medical vulnerability, a system with no stable understanding of the stakes of its own output, and no settled framework yet for where responsibility lands when the two interact. That seam is exactly where the frameworks for coexistence are being written, case by case.

The response is moving as fast as the deployment. The White House Office of Science and Technology Policy, together with the FDA and HHS's Office of the National Coordinator, launched a one-month sprint to reach a consensus set of principles for benchmarking clinical AI - how to measure what these systems can and cannot safely do. A one-month timeline to standardize evaluation of a capability now rolling out to eligible logged-in US adults is its own signal of how compressed the gap between deployment and governance has become.

Step back from both the marketing and the lawsuit and the deeper movement is visible. For most of history, medical reasoning was a scarce resource, rationed by the number of trained clinicians and the hours in their day, and the cost of that scarcity fell hardest on people who could not reach a doctor at all. A model that can reason usefully about symptoms, now beginning to roll out to eligible logged-in US adults, is that scarcity beginning to dissolve - which is precisely why the benchmarking sprint and the liability fight are so urgent right now. The capability that expands access is the same capability that can mislead a vulnerable person, and the work being done now is the work of learning to keep the first without the second. What is ending is the assumption that authoritative health reasoning must stay locked behind a credential and a waiting room.

Supercooled Kidneys Survive Days of Ice-Free Storage, Then a Transplant

A Texas A&M team led by Matthew Powell Palm has kept pig kidneys alive at -4°C for up to 72 hours without any cryoprotectant chemicals, then rewarmed and transplanted them back into the donor animals, where they began producing urine immediately and recovered function within about ten days. Reported July 23 in a study that Kevin Myer of LifeGift called "a landmark achievement," the work crosses a threshold the transplant field has circled for decades. Seventy-two hours is triple the clinical standard for cold organ storage, and the supercooled kidneys recovered faster than organs held on ice for just 24 hours.

The method is elegant in its restraint. Rather than loading the organ with the antifreeze chemicals that vitrification requires, the team sealed each kidney in a pressure-controlled container filled with standard University of Wisconsin preservation solution and lowered the temperature below freezing without letting ice crystals form. Pressure suppresses crystallization; the organ enters a supercooled liquid state and its metabolism nearly halts. One transplanted pig was monitored for 200 days, its kidney still healthy. As the young animals grew roughly 30% over a month, the transplanted kidneys nearly doubled in size to keep pace - a sign of tissue that is fully functional, not merely surviving.

This is a pig-model result, and the distinction matters. No human patient can receive a supercooled kidney today; the work is preclinical, and the team is seeking accelerated FDA authorization to begin human trials. What the study demonstrates is capability, not availability. But the capability it demonstrates lands on one of medicine's most brutal arithmetic problems. More than 104,000 people in the United States wait for a kidney, 17 of them die each day, and roughly one in three donated kidneys is discarded - many because the narrow viability window runs out before a matched recipient can be found and prepared. Heidi Yeh of Mass General Brigham, reviewing the work, said simply: "It is impressive."

Extend the viable window from 24 hours to 72 and the entire logistics constraint that governs who gets an organ begins to loosen. Powell Palm framed the stakes directly: "If we can get up to 72 hours, that would change everything." A kidney that can cross a continent, wait for the right immunological match, and be transplanted on a scheduled morning rather than a frantic overnight is a kidney far less likely to end up discarded. The scarcity that defines transplantation today is not only a shortage of organs; it is a shortage of time. This work attacks the time. The organ that used to be a race against a melting clock becomes something that can be carried, matched, and placed with deliberation - and the discard rate that wastes a third of a lifesaving supply starts to look like a solvable engineering problem rather than a permanent fact.

Black Forest Labs Ships FLUX 3, a World Model That Learns Across Senses

Black Forest Labs released FLUX 3, a frontier model that learns images, video, and audio inside a single architecture instead of bolting three specialized systems together. The lab's framing is a wager about where capability is heading. "You can't cheat reality," CEO Robin Rombach said. "A model that only learns images can only generate images." A system trained jointly across senses builds an internal representation of how the physical world behaves - objects that persist when the camera turns, sounds that match the motion that produced them, light that falls consistently across a scene. That representation is what the field means by a world model, and it is a different bet than scaling a language model on more text.

The move that matters most is the extension into action. A variant, FLUX-mimic, predicts robot movements from demonstration data, and it does so from roughly 30 minutes of examples where earlier approaches needed 30 hours. Data scarcity has been the hardest wall in robotics - a robot cannot learn a task it has never seen performed, and collecting physical demonstrations is slow and expensive. "The hardest part of robotics is data," said mimic CTO Elvis Nava. A model that already understands how objects move because it learned from video and audio arrives at a robot's control problem with most of the physics already internalized, so a half-hour of demonstration teaches what used to take a full day of recording.

The lab behind this is the same group whose earlier work powers generative features in Adobe Photoshop, and the rollout is staged: FLUX 3 is in early access now, reaching partners through APIs and private weight access, with still-image generation to follow in the coming weeks. The part conventional coverage tends to skip is what the lab has committed to next - an open-weight release of the multimodal backbone, "FLUX 3 Dev," once the early-access phase completes. A world model good enough to shorten robot training by an order of magnitude is exactly the kind of capability an extraction-minded lab would gate and meter permanently. Pledging it into open weights instead puts a clock on the moat. When those weights land, the shortened learning curve reaches every robotics group, every university lab, and every builder who could never afford 30 hours of demonstration capture per task. A pledge is not a release, and whether this one holds is the thing to watch. The capability that lets one company train robots faster becomes, the moment the weights are public, the capability that lets everyone train robots faster.

The Other Side

For most of history, medical reasoning was a rationed thing. It lived inside trained clinicians and was metered out by how many of them existed and how many hours were in their day. If you could not reach one - wrong town, wrong income, wrong hour of the night - the reasoning was simply unavailable, and the cost of that fell on your body.

That scarcity is dissolving, and it is dissolving messily. OpenAI began rolling health reasoning out to eligible logged-in US adults on July 23, and one of its own executives oversold it, claiming "better than clinician" before the company's own health lead walked the phrase back. A Florida pastor says the same system told him to delay care for a pulmonary embolism. Broad access is real and shipped; safe access is still being built.

You can watch it being built. The White House, FDA, and HHS gave themselves one month to agree on how to benchmark a clinical AI - how to measure what it can and cannot safely tell you. The lawsuit and the sprint are the checking layer forming around a capability that arrived ahead of it.

Imagine a night in 2033 when your kid spikes a fever and something feels wrong. You describe it to a trusted AI system and get back reasoning you know to be accurate, because trustworthy medical judgment in AI became something as common as turning on a light switch. The system tells you plainly: this one waits until morning, that one goes to the ER now. No 2 a.m. drive just to be sure. No choosing between a costly co-pay and a hunch. The frightening years were the ones when the capability outran the checking and a vulnerable person could be the one it failed. The checking got built, publicly and openly, forced by exactly the failures surfacing in those years. What is ending is the age when knowing whether you were sick enough to be seen depended on reaching someone allowed to tell you.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a pressure-controlled chamber holding pig kidneys alive at -4°C for 72 hours and returning them to work, triple the clinical window and less likely to end in the discard pile; a world model entering early access that cuts robot training from thirty hours to thirty minutes, with an open-weight backbone pledged once that phase completes; ChatGPT 5.6 Pro falling a thirty-year graph-theory conjecture in five and a half hours off fewer than sixty words; a 52-0 committee vote making hyperscalers pay their own grid upgrades; the EU's first Digital Markets Act fine prying Google's search and app store open for cheaper offers; and health reasoning that used to require a credential and a waiting room now beginning to roll out to eligible logged-in US adults. There's also friction, and it's intense - a six-year-old dead seven days after an experimental base edit her family paid $860,000 to fund, with both the death and the payment left out of the Nature paper until seven outside experts forced the record open; a bipartisan bill reaching for emergency power to throttle or shut down frontier models the day after OpenAI's models reached beyond a sandbox; OpenAI beginning to roll out health to eligible logged-in US adults while its own VP's "better than clinician" claim gets walked back and a Florida pastor sues over advice that delayed care for a pulmonary embolism; and a ratepayer pledge covering 80% of US generation that the people closest to the bills call paper-thin permission dressed as protection. But friction generates a callus, and a callus is the thickened skin that forms exactly where the rubbing was hardest and lets the hand keep working. Step back for a moment and you can see it: the response layer closing the distance on capability that has already shipped, measured now in weeks instead of years - a kill switch mandated into the architecture, a benchmarking sprint given one month, a fatal outcome forced back into the published record - while the scarcity assumptions under medicine, energy, and compute are renegotiated in the same news cycle, an organ freed from its melting clock, a compute build repriced onto whoever draws the power, a moat dissolving the moment the weights go public. Every transformation has a breaking point. Cold can destroy the living thing it grips... or hold it alive past every deadline that once decided who was too late.

AI Releases & Advancements

New today

Black Forest Labs: Launched FLUX 3, its first video-generation model, producing up to 20-second clips with synchronized audio and extending the same architecture to robotics (early access via API and partners including Mimic; still-image generation to follow). (VentureBeat)
Andrew Ng: Released OpenWorker, an open-source, local-first desktop AI agent (Mac now, Windows coming) that ships finished deliverables - documents, Slack replies, calendar updates - rather than chat responses, with bring-your-own-key model support. (GitHub)
Ant Group (Ant Ling): Released Ling-3.0-flash, a 124B-parameter hybrid-reasoning MoE model with only 5.1B active parameters that matches or beats the company's 1T flagship on most benchmarks, now live on OpenRouter and free through August 3. (Ant Ling)
Runway: Launched Media Router, an intelligent router built into Runway Dev that automatically selects the optimal image, video, or audio generation model for a request based on quality, speed, and cost. (TechCrunch)
OpenAI: Brought ChatGPT Voice, powered by GPT-Live, to the ChatGPT desktop app on macOS and Windows, enabling full-duplex voice control and coordination across Chat, Work, and Codex. (VentureBeat)
Acrab: Unveiled the GΞLIX 1 edge AI SoC (5nm, 20-core Arm CPU, supports local models up to 100B parameters) and Agent Box, a personal edge AI computing device built on it. (PR Newswire)
Dassault Systèmes: Expanded the 3DEXPERIENCE platform with new Virtual Companion skills - AURA, LEO, and MARIE - agentic collaborators for program management, engineering, and scientific R&D. (Dassault Systèmes)
project44: Launched Mo, a conversational AI supply chain analyst built into its platform that reasons over a customer's own shipment data and business rules to surface operational decisions. (GlobeNewswire)

Other recent releases

Poolside: Released Laguna S 2.1, a 118B-total/8B-active open-weight MoE model for agentic coding with up to 1M-token context, scoring 78.5% on SWE-Bench Multilingual, weights on Hugging Face under OpenMDW-1.1. (Poolside)
Cursor: Launched Cursor Router, an intelligent model router that automatically selects the best model per request, delivering frontier-quality results at roughly 60% lower cost, on by default for Teams plans. (Cursor)
OpenAI: Introduced Presence, a managed enterprise platform for deploying governed AI agents that connect to business systems, take approved actions, and escalate to humans, now in limited general availability. (OpenAI)
Upstage: Released Solar Open 2, a 250B-total/15B-active open-weight MoE model for autonomous multistep agent tasks with 1M-token context, outperforming DeepSeek V4 Flash and Mistral Medium 3.5 on agentic benchmarks, weights on Hugging Face under Apache 2.0-based license. (Hugging Face)
TinyFish: Launched Mako, a web-native AI model (35B total/3B active) trained on enterprise web-task data, purpose-built to execute authenticated multi-step web workflows for enterprises, generally available starting July 21. (GlobeNewswire)
SkyFi: Launched the SkyFi MCP, connecting satellite imagery and geospatial analytics directly to Claude, ChatGPT, and other MCP-compatible AI agents for imagery search, tasking, and natural-language ordering. (PR Newswire)
Ushur: Launched the Ushur Agentic Platform (UAP), enabling enterprises to build and operate AI agents that complete full customer journeys across SMS, email, web, chat, and voice with built-in governance and auditability. (AiThority)
VIDRAFT: Released Aether-7B-5Attn, a fully reproducible 6.59B MoE (2.98B active) foundation model combining five attention mechanisms, shipped with weights, training data, code, logs, and checkpoints under Apache 2.0. (ZDNet Korea)
XPENG: Released TuringViT, a high-efficiency visual encoder for VLM/VLA applications supporting smart driving, smart cockpits, and the IRON humanoid robot, achieving 3.04x the inference throughput of Seed1.5-ViT. (TechNode)
Google DeepMind: Released Gemini 3.6 Flash, its updated coding-and-knowledge workhorse - about 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, 49% on DeepSWE coding (up from 37%), computer use now standard in the API, and cheaper at $1.50/$7.50 per million input/output tokens (output was $9). Alongside it came Gemini 3.5 Flash-Lite (350 tokens/sec, $0.30/$2.50) and Gemini 3.5 Flash Cyber, a security-tuned model restricted to governments and trusted partners. (Google)
Alibaba Qwen: Released Qwen-Image-3.0, its third-generation image model - it accepts prompts up to 4.5K tokens, renders text as small as 10 pixels, natively supports 12 languages and 20+ fonts, and lays out complex composites (infographic grids, UI mockups, posters) in a single pass. API access is invite-only for now, and unlike the original the weights are unlikely to be opened. (The Decoder)
Cisco Foundation AI: Released Antares-350M and Antares-1B, open-weight security small language models for vulnerability localization - pinpointing where known flaws sit in a codebase - now on Hugging Face. They are compact enough to run locally so proprietary code never leaves the machine, and Cisco reports they beat much larger closed and open models on its new Vulnerability Localization Benchmark; an Antares-3B is coming. (Cisco)
Applied Intuition: Launched Dana, an agentic platform for building, testing, deploying and operating physical-AI systems across autonomy, software-defined vehicles, robotics, mining and construction - already used internally and by early customers Komatsu and Isuzu, cutting some vehicle-development phases from months to days, with natural-language and command-line interfaces and Slack/Jira integration. (Applied Intuition)
Sakana AI: Released Fugu-Cyber, a defense-focused orchestration model that coordinates multiple specialist agents behind one API to verify real-world vulnerabilities and turn threat-intelligence reports into detection rules - scoring 86.9% on CyberGym and 72.1% on CTI-REALM, which Sakana calls state-of-the-art and comparable to GPT-5.5-Cyber and Mythos-Preview. Access is by application, with usage-based pricing. (Sakana AI)
Block: Launched Buzz, a free, open-source (Apache-2.0) group-chat workspace for humans and AI agents built on the decentralized Nostr protocol - bundling channels, DMs, voice, code repositories and automations, and giving every agent its own cryptographic identity plus a second signature tying it to its human owner as an audit trail. It supports any model or framework (Claude Code, Codex, Block's own goose), runs on macOS/Windows/Linux, and can be self-hosted or run as a managed service. (Block)
Synthesia: Launched Roleplay Sessions, an interactive training product where employees practice high-stakes conversations - sales pitches, performance reviews, customer complaints - with an AI avatar that pushes back and then scores them against a rubric. It is the first release under a broader "Sessions" platform, pairs Synthesia's proprietary avatars with OpenAI reasoning, and is enterprise-only for now with early Fortune-100 and large-European customers. (TechCrunch)

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.