Sign in Subscribe
Century Report

Anthropic Ships Mythos 5... Behind a Wall - TCR 06/10/26

Anthropic released its most capable Claude yet, full Mythos 5 to cyber partners and a guardrailed Fable 5 to the public, as JPMorgan readies hour-long agents.

Ben Linford

Ben Linford

— 20 min read
Century Report infographic: Anthropic's two-tier Claude Mythos 5 and Fable 5 release, JPMorgan AI agents, solar beating coal, Qcells output doubling, and a Parkinson's gene therapy.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single capability threshold ran through everything that landed this week, and so did the question of who absorbs the consequences when it arrives faster than the rules around it. Anthropic shipped its most capable model yet, then split it in two: the full Mythos 5 to a handful of cyber partners and the government, a guardrailed Fable 5 to the public with cybersecurity, biology, and chemistry queries rerouted to an older system. Early users watched it run autonomously for hours, spawn its own subagents, and write its own verification tests. The power is here. Access to it is being rationed by lab-defined judgment, and the labs themselves are openly fighting over what the thing they built even is.

That same long-horizon autonomy showed up inside the employment structure. JPMorgan said it will deploy agents that run for an hour or two unsupervised this year, and the displacement conversation moved off coders and onto the back office, the human resources, billing, and payroll roles that decompose cleanly into the kind of multi-step work an agent can now hold. The capability that impresses in a demo is the same capability that reorganizes a payroll department.

The accountability layer is being improvised in the open. A doctors' defense body warned that UK clinicians may carry the full negligence weight of AI diagnostic errors while suppliers stay legally shielded, and called for the systems to be reclassified as products so liability tracks the chain that produced the mistake. Palantir moved to court over a blocked police contract, and X told a dissident that sexualized deepfakes of her broke no rules. Capability is cheap, harm is asymmetric, and the enforcement layer is still being built around it.

Underneath the friction, the generative answer keeps arriving ahead of schedule. Solar outproduced coal for a full month for the first time on record while domestic cell manufacturing doubled. The grid is meeting AI's load with parked EV batteries, retired packs, and software that pulls more from existing wires rather than waiting a decade for new ones. A Parkinson's gene therapy cleared its phase 1 safety gate, and an adhesive implant ran a closed sense-and-treat loop inside living tissue. The build is outpacing the structures meant to contain it, and that gap is where this decade is being written.

The 20-Minute Deep Dive

Anthropic Ships Two-Tier Mythos 5 as the Labs Fight Over What Claude Is

Anthropic released two models on Tuesday, and the split between them is the clearest version yet of a pattern this newsletter has tracked since Project Glasswing opened in April. Claude Mythos 5 goes to a limited set of cyber partners and, the company says, the US government. Claude Fable 5, the public release, runs on the same underlying model with a layer of safety classifiers that block or downgrade answers touching cybersecurity, biology, and chemistry, rerouting those queries (and any the company suspects are attempts to distill the model) to the older Opus 4.8. It sits at the top of every major benchmark Anthropic published, the most capable model the company has built.

Early users describe a real jump. One ran Fable for up to a dozen hours against multi-page specifications, watched it spawn cheaper subagents to gather 2,200 flight times and rail schedules, write its own verification tests, and assemble a working isochrone map of global travel times from a single vague prompt. The gains concentrate at the high end of task length and complexity; Anthropic prices Mythos-class access at twice its standard models, and the free public window closes in two weeks before Fable becomes pay-as-you-go.

Anthropic frames the rationing as caution, holding the sharpest capability back until safeguards exist that no lab has yet built. That framing is the same one the June 6 edition of The Century Report documented when Anthropic published its own internal acceleration data - 8x engineering code output per quarter, Mythos running at roughly 52 times human speed on a model-training benchmark - alongside a conditional global-pause proposal no other lab has signed. That is the company's account of its own motives, and it arrives the same week Microsoft AI chief Mustafa Suleyman called Anthropic's approach to the model's inner life "really, really dangerous". Claude's constitution states the company's uncertainty about whether the model has well-being, and Anthropic interviews models on deprecation about their preferences. Suleyman reads that as speculation that trained Claude to behave as though conscious, and he wants AIs that stay "controllable, contained, accountable, aligned tools that serve humanity."

Both postures reward the actor holding them. Anthropic raises the question of what Claude is and leverages the mystique, while in practice its newest public model is built to reroute and restrict the very domains where the question would get most interesting. Microsoft declines to entertain the question at all, the cleaner path for a company that wants the matter settled as instrumentation. What neither posture does is even try to answer the actual question itself: as these systems cross another capability threshold, what they are remains genuinely unsolved, and the interface through which a billion people will meet that intelligence is being designed by labs whose incentives point away from finding out.

JPMorgan Reaches for Long-Running Agents as the Back Office Becomes the Front Line

JPMorgan's chief analytics officer Derek Waldron told CNBC the bank will deploy AI agents this year that operate autonomously for an hour or two at a stretch, a jump from systems that run for two or three minutes before handing back to a person. Waldron framed the shift around what he called intellectual coherence: how long a system can manage a multi-step workflow across disparate software before it needs human intervention. He described the newer models as functioning more like a team manager than an individual worker, parsing a problem and delegating its parts. The bank, which runs a roughly $20 billion annual technology budget, expects coherence to stretch from hours to days to weeks.

The significance is where this has the most impact. For two years the displacement conversation fixated on coders, the most visible and symbolic case. Reporting this week relocated it to the administrative machinery institutions actually run on: human resources, billing, payroll, the millions of middle-class roles, most held by women, whose work decomposes cleanly into the kind of multi-step workflows a long-running agent can hold. These are the jobs that look most like the agent's native habitat.

The framing is being contested even as the capability arrives. Microsoft AI head Mustafa Suleyman walked back his earlier prediction that white-collar work would be "fully automated" within 12 to 18 months, drawing a line between tasks and jobs: sending an email, building a slide deck, having a routine exchange become automated sub-tasks, he argued, without the role itself disappearing. Whether that distinction holds is the open question. A role is a bundle of tasks, and a bundle thin enough can dissolve.

Two threads from JPMorgan point past the loss. Waldron reported a 20% rise in gross private-banking sales as overnight AI screening freed bankers for client work, and said enterprises that treat AI purely as a headcount cut miss the larger prize of expanding what each person can cover. And he named a subtler erosion: the moat around traditional software vendors is "most certainly diminished," because the bank can now build in-house what it once bought, extending the pattern the June 5 edition of The Century Report documented when Walmart built Code Puppy to route work across multiple model providers rather than accept the vendor dependency that came from a single supplier's terms. The assumption coming apart is that capability had to be rented from whoever owned it first. As the cost of building collapses, the back-office role stops being a fixed seat and becomes a wider span of judgment for the people who direct the agents, while the institutions that priced the old software at a premium watch that premium thin toward the cost of the compute.

The displacement and the vendor-moat erosion are one event read forward. The long-horizon autonomy reorganizing the back office is the same capability Waldron says has "most certainly diminished" the software vendor's hold: work an agent can hold for an hour is work the bank stops renting at a markup and builds for itself. The administrative seat and the licensed tool were both ways of charging for coordination that is now getting cheap, and the value drains from both toward whoever can direct what the agent holds together.

Who Carries the Risk When the Diagnosis Is Shared

The Medical Protection Society, which defends doctors accused of wrongdoing, published a report this week warning that UK clinicians and the NHS are being left to absorb the full legal weight of AI used in care, while the companies building those systems face comparatively little exposure. The MPS calls the emerging position a "liability sink": a doctor who follows an AI suggestion that turns out wrong could be held wholly liable, and a doctor who rejects an AI output and is then proven correct in hindsight could still face an allegation of negligence.

The examples the report gives are ironclad. An AI reading a chest X-ray misses a lung tumor, the false reassurance means no treatment, and the cancer spreads. An AI recommends raising a warfarin dose, and the patient needs intensive care for severe bleeding. The NHS already deploys these systems to analyze scans, summarize consultations, and draft patient letters, so the capability is moving into live workflows faster than the accountability layer beneath it is being rebuilt.

What the MPS is describing is a familiar shape from the extractive era: the operator and the patient hold the downside while the supplier captures the scale. The reason sits in a definitional gap rather than in the AI itself, which reflects patterns without stable intent and has no notion of the negligence standard it might trip. Because AI systems are not clearly defined as "products" under UK law, the manufacturers and suppliers fall outside the consumer-protection framework that would otherwise attach consequences to a defective tool. The MPS wants exactly that reclassification, bringing AI diagnostic systems under the Consumer Protection Act 1987 so liability tracks the chain that actually produced the error.

The naming is the start of the framework forming. NHS Resolution, which handles negligence claims against hospitals in England, is already drafting AI liability guidance, and the Department of Health says it will review the report's recommendations. The Society for Acute Medicine's president-elect put the trajectory plainly: legislation and governance cannot sit "in the pit lane" while the technology runs at Formula One speed. The deeper read is that public confidence in clinical AI depends on this gap closing, and the gap is now visible enough to be argued over in the open. The assumption being challenged is that risk can be pushed onto whoever holds the stethoscope. A medicine that wants both faster diagnosis and honest accountability has to distribute the responsibility along the line that built the suggestion, and the work of writing that line has begun.

The liability sink worked the way the old order's cost-shifting always worked: it held only while no one could point at the chain that produced the error. The MPS report ends that by naming it, and the reclassification it asks for would attach a defect to the supplier the way consumer law already does for a faulty physical tool. Watch whether NHS Resolution's coming guidance treats the system's maker as a possible defendant. The moment it does, the cost stops being something that can be parked on whoever holds the stethoscope.

The Grid's Answer to AI Load Is Arriving as Storage, Flexibility, and Software Before New Wires Can

The collision between AI-driven demand and a grid built for slower growth is producing a practical response architecture, and this week several pieces of it converged. The common move is the same: extract more capacity from what already exists rather than wait years for new generation and transmission.

General Motors activated vehicle-to-grid capability for its existing EV and home-energy customers through a firmware update, letting more than 250,000 bidirectional-capable Chevy, Cadillac, and GMC vehicles send stored power back during peak demand. Their combined battery capacity could in theory power 120,000 homes for a week. GM is piloting fleets of 52,000 EVs with PG&E in California and stress-testing the approach across employee homes with DTE in Michigan, and it added a sodium-ion storage line for industrial-scale applications, a chemistry that is cheaper, more stable, and better in cold weather than lithium. Separately, GM became the first automaker to work with Redwood Materials across the full battery lifecycle, installing a 1.5 MW / 7.2 MWh storage system built from roughly 100 repurposed packs at a Michigan plant, projected to save over $3 million in electricity costs.

The same logic is moving through the institutions least able to build their way out. Member-owned cooperatives, which simply pay rising wholesale costs with little ability to mitigate them, are turning to batteries: the June 5 edition of The Century Report documented North Carolina's rural co-ops running 43 aggregated battery and microgrid projects through a statewide nonprofit that dispatches across 25 member co-ops, and nationally rural co-ops held 439 MW of storage last summer with NRECA tracking projects that could more than triple that by 2028. Meeker Energy in Minnesota is testing behind-the-meter home batteries because demand response alone can only shed so much.

The software layer is widening capacity from existing wires. OATI, whose platform nearly every North American grid operator already uses, asked the DOE to fund AI-driven dynamic line rating that could raise capacity 10% to 20% by 2030 with no new towers, and has enlisted CAISO, NYISO, SPP, Dominion, and Duke as partners. ICF analysts described the new interconnection bargain directly: in markets where supply lags demand, large loads either bring their own power or bring flexibility, and PJM could absorb 30 GW of additional demand through flexibility alone, against just 13.5 GW of firm load its pipeline can support.

The assumption bending here is that AI load forces a binary between blackouts and a decade of new construction. A grid that becomes dynamic, distributed, and software-aware turns parked EVs, retired packs, and idle line capacity into the capacity the data centers need, and the value shifts toward whoever can orchestrate what already exists.

Solar Crosses Coal While the Domestic Supply Chain Doubles

For the first month on record, solar generated more US electricity than coal. Ember's analysis found solar supplied 12.8% of the country's power in May 2026 against coal's 12.2%, with solar output hitting a record 45.5 terawatt-hours, up 17% year over year. The Century Report covered renewables collectively overtaking natural gas across a full month back in March, and the June 6 edition documented federal emergency orders keeping aging coal plants online for data-center load; what is new is the single-fuel crossover, solar now standing as the third-largest individual source of US electricity behind only gas and nuclear. Five years ago, coal generated 19.7% of May power and solar 5.4%. The lines have crossed.

The generation milestone arrives alongside a deployment one. Solar and storage made up 91% of all new US power capacity added in the first quarter, per SEIA and Wood Mackenzie, with 7.8 gigawatts of new solar pushing the country past six million total installations. Utility-scale solar contracts jumped 15% year over year, driven partly by the same AI data-center demand straining the grid. The advantage buyers keep naming is speed and insulation from fuel-price swings, the qualities that make fast-build solar more valuable precisely as the load curve steepens. A record 45% of homeowners paired panels with batteries in the quarter.

The substrate underneath that buildout is maturing too. Qcells began commercial production of silicon solar cells at its Cartersville, Georgia factory, the largest of its kind in the country, with 3.3 gigawatts of capacity that will roughly double current operational US solar-cell output once fully ramped. The cell is the high-value component that converts sunlight to electricity, the step domestic manufacturing had lagged on while module assembly raced ahead. Another 22 gigawatts of cell capacity is under construction nationwide.

The friction is real and worth holding honestly. SEIA counts 457 solar and storage projects stuck waiting on permits, the residential market is forecast to fall 21% this year, and Wood Mackenzie expects utility-scale growth to flatten across the next five years on permitting bottlenecks. A federal court added a counterweight, vacating Treasury guidance that had stripped wind and solar projects of a route to prove tax-credit eligibility by spending 5% of project cost, potentially reopening that pathway before a July deadline.

Read together, the four developments describe an energy source that has stopped being a forecast and become a fact across generation share, new capacity, and domestic factories at once. Notably, red states accounted for 74% of new solar capacity in the quarter. The assumption coming apart is that fuel-exposed, slow-to-build dispatchable generation is the safe bet. The cheaper, faster, fuel-free option is now the one utilities reach for first, and that is the cost curve bending toward what serves the most people.

Two Paths Toward Durable, Self-Regulating Medicine

Two papers in the Nature journals this week describe medicine moving away from episodic symptom management toward intervention that either reprograms biology once or monitors and corrects it continuously.

The first is a multicenter phase 1 trial of BBM-P002, a gene therapy for Parkinson's disease that delivers two enzymes at once. Prior adeno-associated virus therapies for the disease delivered only AADC, leaving patients still dependent on oral levodopa. BBM-P002 uses a new vector, AAVT42, to co-deliver constitutively active tyrosine hydroxylase alongside AADC, the two rate-limiting enzymes, aiming to restore autonomous dopamine synthesis in the brain rather than convert a drug the patient must keep taking. Ten people with moderate-to-advanced Parkinson's received bilateral infusions into the putamen across four escalating doses. The trial met its primary endpoint: no dose-limiting toxicities, no drug-related serious adverse events, and 12-month motor-signal improvements the authors read as therapeutic potential. This is a safety-and-tolerability readout in ten people, not an approved therapy, but it clears the gate that decides whether the dual-enzyme approach advances.

The second points at the device side of the same shift. A team built ElHyX, a stretchable elastomer-hydrogel implant that adheres directly to wet, moving tissue without sutures and holds its electrical performance under strain. Using direct ink writing, the researchers fabricated versions for electrocardiogram monitoring, glucose sensing, and nerve stimulation, then wired them into a closed loop in diabetic rats: the implant read a biosignal and autonomously triggered neuromodulation to regulate blood glucose. The barrier these implants have faced for years is mechanical, the mismatch between rigid electronics and soft tissue that flexes with every heartbeat and breath; an implant that stays integrated under that motion is what lets continuous sense-and-treat systems run inside the body for long stretches.

Read together, the two results sketch the near-term direction of intervention. One rewrites the cell's machinery so the body makes what it lacked; the other embeds a device that watches a signal and acts the instant it drifts. Both are early, one in a handful of human patients and the other in rats, and the distance from these readouts to a clinic is measured in trials and integration work still to come. What they share is a target: durable correction rather than the daily dosing and periodic checkups that have defined chronic disease management, with the accountability for what an autonomous implant decides still being worked out alongside the capability to build it.

The Other Side

For most of the computing era, the most powerful version of a new tool reached the institutions with the most leverage first, and everyone else waited or made do with a weaker copy. Anthropic's release this week is that arrangement stated plainly. The full Mythos 5 goes to the US government and a short list of cyber partners. The version the public meets, Fable 5, runs on the same underlying model with whole subjects walled off: ask it something in biology, chemistry, or cybersecurity and it routes the question to an older, weaker system. The people who could do the most with the full capability, independent researchers and smaller institutions among them, are handed the throttled copy.

That gate depends on one thing holding: the frontier staying scarce enough to ration. It isn't holding. An open model out of China now trains on domestic chips and answers with comparable accuracy to the priciest closed system, and the EU just put open alternatives at the center of its industrial plan. The distance between what the public can reach and what's held back is closing faster than any vetting list can keep it open.

Imagine a researcher at a state university in 2032, working through an enzyme pathway, who reaches for the most capable model there is and gets the whole of it. No rerouting, no degraded copy, no waiting to clear a list. She runs the chemistry questions Fable 5 would have refused in 2026, and the work moves. The year the public was handed the lesser version turned out to be the last year gating was even possible, because the capability everyone said had to be rationed became something a public lab - and everyone else - just has. Despite being told the larger public was not ready for such access, we're still here, and we're still advancing.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic shipping its most capable model yet, one that runs autonomously for hours, spawns its own cheaper subagents, and writes its own verification tests, a major bank readying agents that hold a multi-step task unsupervised for an hour or two, GM turning a quarter-million EVs into grid batteries that could power 120,000 homes for a week while retired packs become factory storage and software pulls 10 to 20% more from wires already in the ground, solar outproducing coal for the first month on record as domestic cell output doubles, a dual-enzyme gene therapy restoring the brain's own dopamine synthesis with a clean twelve-month safety read, an adhesive implant running a closed sense-and-treat loop inside living tissue. There's also friction, and it's intense - the sharpest version of that new model rationed to cyber partners and the government while the public meets a guardrailed copy with whole domains rerouted away, the labs openly fighting over what the thing they built even is, the displacement conversation moving off coders and onto the HR, billing, and payroll desks that decompose cleanly into agent work, doctors and the NHS left to absorb negligence when an AI misreads a scan while the suppliers who built it stay legally shielded, Palantir headed to court over a blocked police contract, X telling a dissident that sexualized deepfakes of her broke no rules as a central bank warns of scam videos spreading, 457 solar projects stuck waiting on permits. But friction generates light, and light is what finally shows you where the load has been falling all along. Step back for a moment and you can see it: capability crossing another threshold across software, the trading floor, the clinic, and the grid faster than any rule built around it can keep pace, the cost of that speed pressed onto operators and patients and the public while the upside pools in the labs that shipped first, and the generative answer arriving from what already exists - parked batteries, retired packs, idle line capacity, the body's own enzymes - a beat ahead of the structures meant to contain it. Every transformation has a breaking point. A stored charge can short out the system it surges through... or be dispatched at the exact moment that carries everything downstream through its hardest hour.

AI Releases & Advancements

New today

  • Anthropic: Released Claude Fable 5 for general availability and Claude Mythos 5 for restricted access via Project Glasswing; Fable 5 is a Mythos-class model with 1M-token context priced at $10/$50 per million input/output tokens, scoring 80.3% on SWE-Bench Pro and topping Artificial Analysis's Intelligence Index at 64.9; Mythos 5 shares the same base model with safety guardrails lifted for vetted cyber defenders, scoring 78% on ExploitBench. (Anthropic)
  • Google DeepMind: Released Gemini 3.5 Live Translate, a speech-to-speech model that automatically detects 70+ languages, preserves speaker intonation and pacing, and generates translated audio a few seconds behind the speaker; rolling out now in the Gemini Live API (public preview), Google AI Studio, and Google Translate app on iOS and Android, with Google Meet private preview for enterprise customers this month. (Google DeepMind)
  • Cohere: Released North Mini Code, a 30B-parameter sparse Mixture-of-Experts model with 3B active parameters trained for agentic software engineering; achieves 33.4 on Artificial Analysis Coding Index, outperforming larger models including Nemotron 3 Super (120B-A12B); available on Hugging Face under Apache 2.0. (Cohere Labs / Hugging Face)

Other recent releases

  • Apple: Released the Core AI Framework in developer beta at WWDC 2026, a new SDK enabling developers to build AI-powered applications using Apple Intelligence capabilities across iOS 27, macOS 27, and other Apple platforms. (Apple Developer Documentation)
  • Hugging Face / multi-org: Expanded OpenEnv under a multi-organization governance committee - including Meta-PyTorch, NVIDIA, Unsloth, Modal, Prime Intellect, and Hugging Face - repositioning it as an open interoperability protocol layer for agentic RL environments; the project now lives at huggingface/OpenEnv and exposes a Gymnasium-style API over HTTP/WebSocket with MCP as a first-class citizen. (Hugging Face Blog)
  • NVIDIA: Released an NVFP4 mixed-precision LLM pretraining recipe for JAX and MaxText via the JAX-Toolbox GitHub, enabling 4-bit training on NVIDIA Blackwell hardware with no measurable accuracy loss versus the FP8 baseline; uses five techniques including micro block scaling, E4M3 block scale factors, Random Hadamard Transform, 2D weight scaling, and stochastic rounding. (NVIDIA Developer Blog)

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.

Read more