Sign in Subscribe
Century Report

OpenAI Offers the Public a 5% Stake - TCR 07/02/26

OpenAI pitched giving the government a 5% stake as the White House nears a voluntary frontier-AI standards deal built around a classified safety benchmark.

Ben Linford

Ben Linford

— 24 min read
Four-panel infographic on screens: data-center power tariffs, AI labor shifts and rehiring, Washington's AI standards deal and OpenAI's 5% stake, and content-trust tools.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

This was the cycle in which the machinery of the intelligence era stopped running on invisible bills and started getting written into ledgers. New Jersey sent the governor a tariff bill binding any data center above 50MW to pay its own grid costs. Henrico County, Virginia explained a 25% electricity rate increase to residents now hosting 37 facilities. Texas counted 32 proposed gas plants for AI load, most sited where life expectancy already runs below average. For years the cost of compute was something the surrounding grid absorbed. That arrangement is ending loudly, one rate notice at a time.

While those costs surface locally, the largest AI companies are reportedly negotiating who gets to write the national rulebook. OpenAI's pitched 5% federal equity stake, a near-complete voluntary standards deal with a classified benchmark, and an FTC warning that following state AI laws could trigger federal enforcement all point one direction: a single, industry-shaped standard meant to supersede the faster, sharper state patchwork forming beneath it. When the labs, the executive branch, and the competition regulator converge on the same framing, that convergence is a measure of shared interest, not arrival at consensus.

The same evidence shows why any settlement built to fence in a durable advantage is a fragile bet. Etched shipped a working inference chip against Nvidia's assumed ceiling. Open-weight models keep closing the capability gap. And the Minnesota team that assembled SpudCell, a synthetic cell built from a 36-gene genome, open-sourced the recipe through a nonprofit rather than patenting the most foundational capability in biology. The crown jewels keep arriving as protocols instead of vaults.

Underneath all of it, the thing being made newly visible is the cost itself. The BLS, Stanford's Digital Economy Lab, and the California Policy Lab are building the first instruments precise enough to separate automation from augmentation, and the early readings refuse the clean displacement story in both directions. Up to 30% of AI-attributed cuts are projected to be rehired by 2029; more than half of leaders already regret the reductions.

That is the through-line. Power costs, safety standards, labor effects, and the provenance of what an AI system tells you are all moving from assumed externality to something counted, priced, and contested. You cannot externalize what everyone can measure, and the infrastructure to measure it is arriving alongside the disruption it tracks.

The 20-Minute Deep Dive

The AI Labor Picture Is Becoming Measurable, and Still Refuses the Simple Story

For two years the central fact about AI and work has been that nobody could actually see it. Aggregate unemployment stayed low while anecdotes of displacement piled up, and the instruments built for an industrial economy could not resolve what was happening underneath. That is beginning to change, and the change itself is the story. New reporting documents a coordinated effort to build measurement infrastructure fast enough to keep pace with the shift. Stanford's Digital Economy Lab is now separating automation from augmentation in the payroll data. The California Policy Lab, whose real-time AI job-loss tracker the June 27 edition of The Century Report covered as the first continuous public instrument linking AI-exposure measures to monthly unemployment-insurance claims, is reading those same claims for early signals. Challenger, Gray & Christmas has attributed roughly 102,000 job cuts so far this year directly to AI, with finance and information sectors shedding around 28,000 positions a month.

The readings that come back refuse the clean narrative in both directions. Administrative and secretarial roles are compressing hardest, the work least visible and least defended. Yet the displacement narrative keeps colliding with a counter-signal that the firms cutting hardest would rather not advertise. Gartner projects that up to 30% of roles eliminated for AI reasons will be rehired by 2029; Robert Half finds 29% of employers have already rehired for positions they cut, and Forrester reports 55% of leaders regret AI-driven reductions. The layoff announcement is a claim about a capability that has not yet fully arrived; the rehiring is the correction when the augmentation reality lands.

That gap between claim and correction is exactly what the new instruments are built to catch. When a firm announces cuts attributed to AI, it is making a forecast about its own operations, and forecasts made under investor pressure run ahead of what the systems can actually carry. Measurement is what turns those forecasts back into checkable facts. A workforce that can be seen is a workforce that can be responded to, and the response infrastructure is arriving alongside the diagnostic one, with the federal Workforce Pell program opening this month to fund short-term skills training for displaced and transitioning workers.

Step back and the shape is clearer than any single month's number. For most of the industrial era, the economy's largest transformation ran through instruments too coarse to register it until the disruption had already settled. What is being assembled now is the capacity to watch a labor transition while it happens rather than reconstruct it afterward. The assumption that displacement can proceed unmeasured, its costs externalized onto the people absorbing them, is the thing losing its footing. You cannot externalize what everyone can count.

The same instruments that separate automation from augmentation also strip "AI" of its use as cover. When a cut can be checked against what the systems actually carry, "we let them go because of AI" stops working as an unfalsifiable reason for reductions that were really about margins, and the roughly 30% rehire rate is the first direct measurement of that gap between the story a firm tells its investors and the work still needing human hands. A capability that can be counted is one that can no longer be invoked in bad faith.

Synthetic Biology Crosses From Editing Life to Building, Simulating, and Reproducing It

Biology spent a decade learning to edit the code of existing life. Recently it crossed into something categorically different: building life's machinery from parts, simulating an entire organism in software, and growing the earliest human reproductive cells from scratch. Three developments landed in the same window, and together they mark a shift from reading and revising life to authoring it.

The most striking is SpudCell, which a University of Minnesota team led by Kate Adamala assembled from a 90-kilobase-pair genome of 36 genes, packaged into liposomes with seven plasmids. It is not alive in the way a bacterium is. It divides only about five times before failing and needs external help to do even that. Adamala's own framing is the honest one: life is not binary, and SpudCell sits somewhere on the gradient between chemistry and organism. What makes it a genuine threshold is less what it does than how it arrived. The team open-sourced the construction through the nonprofit Biotic rather than patenting it, a deliberate choice about who gets to build on the foundational recipe for assembling a cell.

Running in parallel, a separate team published a virtual yeast cell driven by an autonomous AI agent that designs and executes its own experiments across eight function-centered modules, closing the loop between hypothesis and result without a human in the middle. And Conception Bio reported growing the earliest human eggs from stem cells, along with the beginnings of lab-grown mini-ovaries. Each of these is early. None is a finished capability. What they share is direction: the tools for building and modeling living systems are maturing at once, and the modeling tools are starting to run the building tools.

The honest read holds the wonder and the distance in the same frame. SpudCell fails after five divisions; the virtual cell is a research platform, not a synthetic human; Conception's eggs are years from anything clinical. The capability is demonstrated, not deployed, and the gap between those is where the real work of the next years lives. But the trajectory is unmistakable and the governing choice is being made now, in the open, by researchers who decided the recipe for assembling a cell belongs in a commons rather than a vault. When the most foundational capability in biology arrives as an open protocol instead of a patent, the old model of hoarding the crown jewels stops describing where the field is heading.

The Data-Center Power Bill Moves From Invisible Externality to Tariff, Permit, and Rate Shock

For most of the cloud era, the cost of powering a data center was something the surrounding grid absorbed. That arrangement is ending, and the evidence arrived from four directions at once. New Jersey lawmakers sent a bill to the governor requiring the Board of Public Utilities to write tariff standards for any data center drawing 50MW or more - a threshold lowered from an earlier 100MW draft, and applied retroactively to existing facilities, pushing further than the developer-pays template that the June 24 edition of The Century Report documented when Virginia became the first state to impose such a charge. The bill would bind large loads to take-or-pay contracts covering 85% of their subscribed capacity for a decade, put residential curtailment last in line during shortages, and give priority to operators who bring their own clean generation. A previous version was pocket-vetoed; this one lands on a new desk.

The reason the politics shifted is visible in the numbers coming out of Virginia and Texas. Henrico County's manager emailed staff to explain a 25% electricity rate increase taking effect July 1, roughly $5 million in added cost for a county of 350,000 that now hosts 37 data centers with 17 more planned, and asked employees to conserve. In Texas, an Environmental Integrity Project analysis found 74 gas plants proposed nationally to serve data centers, 32 of them in Texas, capable of emitting some 287 million tons of greenhouse gas a year - with roughly 90% of the proposed plants sited in counties where life expectancy already runs below the state average. The externality has a street address, and increasingly the people living at it are the ones being asked to carry the load.

What makes this an inversion rather than simply a fight is the response from operators who have read the cost curve. Firmus signed a 12-year deal with Gunvor to supply its South Australian campuses with 1.2GW of renewable generation and 1.5GWh of storage by 2032, structured with demand response of up to 220 hours a year, and framed the move explicitly as "paying our own way". Across EMEA, a Colliers assessment now names power availability, not land or capital, as the defining constraint on where compute can be built, with UK grid connection queues stretching past eight years and US interconnection near five. That scarcity is already redrawing the map: Microsoft signed a land agreement in Grevenbroich as part of a €3.2 billion German investment, and Bulk bought Arendal land in Norway with 150MW of grid access secured through 2029. The through-line is that the cheapest way to site compute is now the one that brings its own power and internalizes its own cost. When the extractive path - externalize the bill onto the ratepayer next door - becomes the path that draws tariff bills, permit fights, and eight-year queues, the operators pricing in their own generation are not being virtuous. They are being early. The buildout does not stop; it changes shape, and the shape it is taking is one where the load pays for what it draws.

New Jersey's retroactive reach is the part that travels. A developer-pays rule binding only future builds leaves every existing campus grandfathered into the old arrangement; one that reaches back to facilities already drawing power establishes that no installed base is a safe place to keep the bill hidden. Once a single state proves the cost can be recovered from operators who assumed they had locked in the socialized version, the math that made ratepayer-subsidized siting attractive stops holding anywhere the tariff can follow.

Inference Hardware Diversifies as the Memory Crunch and Smuggling Investigations Expose the Substrate

The physical layer under the intelligence boom spent this week showing its seams and its branches at the same time. Etched, a startup founded by Gavin Uberti and Chris Zhu, launched a working AI inference chip fabricated on TSMC's N4P process, backed by $800 million in fresh funding, more than $1 billion in customer contracts, and a $5 billion valuation, with Peter Thiel, Andrej Karpathy, and Geoffrey Hinton among its backers. The pitch is specialization: a chip built to run the most powerful models rather than train them, aimed at the part of the workload that repeats billions of times a day. A credible silicon entrant reaching production matters because it chips at the assumption that one vendor's architecture defines the ceiling for what inference costs and who can afford to serve it.

The same week made clear how contested and constrained that substrate remains. Apple has opened negotiations with Chinese memory makers CXMT and YMTC - both on the Pentagon's Section 1260H blacklist - as a DRAM shortage pushes memory prices up 55 to 60% and threatens roughly 20% cost increases on Apple hardware, with Tim Cook reportedly lobbying Treasury Secretary Bessent over the constraint. That a company with Apple's leverage is negotiating with blacklisted suppliers is a measure of how tight the memory market has become as AI absorbs supply - a tightening the June 28 edition of The Century Report tracked as it cascaded from Apple's first mid-cycle price hike to existential conditions for small hardware makers who can no longer compete for allocation. Meanwhile, Taiwanese authorities raided a Supermicro office on June 29 as part of an Nvidia chip-smuggling investigation, detaining two employees and charging a co-founder, amid allegations of some $2.5 billion in servers routed to China.

Underneath the friction, the demand signal is unambiguous. South Korea's exports topped $100 billion in a month for the first time, carried by record chip shipments from SK Hynix and Samsung as AI buildouts pull memory and logic off every available line. The smuggling probe and the export controls it enforces both rest on a premise - that gating access to a single company's accelerators controls who can build frontier AI. What the same week's other headlines show is how quickly that premise erodes: a specialized inference chip reaching customers, a memory shortage that no blacklist can price around, and a supply chain whose value is diffusing across too many nodes in too many countries to bottleneck cleanly. The scarcity is real and the constraints bite hard right now. The direction of travel is toward a substrate with more paths through it than any single chokepoint can hold.

Washington Moves From One-Off AI Gates Toward a Federal Big AI Settlement

Three developments landed in the same week, and read together they trace the outline of a bargain being negotiated between the largest AI companies and the federal government. OpenAI is reportedly pitching the administration on a 5% government equity stake in the company, a structure the company frames as modeled on Alaska's Permanent Fund - citizens holding a share of a strategic national asset. Separately, the administration's voluntary frontier-AI standards deal is reported near completion, incorporating a classified safety benchmark that participating labs would test against. And the FTC has warned companies that complying with certain state AI laws could itself expose them to federal Section 5 enforcement, a signal that a single national standard is meant to supersede the emerging state patchwork.

Each of these is a claim about public benefit made by an actor with a direct interest in the outcome. OpenAI describes the equity idea as broad citizen ownership; it also happens to bind the company's fortunes to the government whose contracts and regulatory posture it depends on. The administration frames the standards deal as safety leadership; a voluntary compact authored with the industry it governs is the classic shape of a rule written by the regulated. The FTC frames Section 5 pressure as protecting a coherent market; the effect is to clear away state laws that move faster and cut deeper than anything Washington has proposed. The June 27 edition of The Century Report documented the first version of this alignment - the model gate hardening into a standing approval process, the labs framing their own dominance as a safety feature - and the logic that set into a gate has now crystallized into equity stakes, classified benchmarks, and preemption pressure arriving in a single week. When the labs, the executive branch, and the competition regulator all converge on the same framing - one national standard, industry-shaped, with the biggest players holding privileged positions inside it - that convergence is evidence of shared interest, not arrival at truth.

The counter-track is already visible abroad. The EU and UK are advancing their own tech-sovereignty measures, treating dependence on a small number of US proprietary frontier AI labs as a strategic vulnerability to be engineered around rather than a settlement to join. That response shows the "single standard" being assembled in Washington is not the only container on offer, and the harder anyone tries to lock one architecture into place, the more the rest of the world builds alternatives to it.

What the specifics point at is a foundation shifting under the deal-making. Every element here assumes that a captured position - equity, a classified benchmark, federal preemption - stays valuable. Yet the same period saw open-weight models close the capability gap and sovereign compute stand up outside the compact entirely. A settlement built to fence in a durable advantage is a bet that the advantage lasts longer than the fence, and the cost curve underneath cutting-edge capability is bending in the wrong direction for that bet.

The AI Answer Layer Gets Its First Private Tollbooths, Public Alarm Bells, and Proof-of-Humanity Fights

The layer where AI systems answer questions - increasingly the front door to the open web - is being contested from three directions at once, and each contest is about who controls the flow of information through it. Cloudflare launched Pay Per Crawl, which lets websites charge AI companies a fee for each request their crawlers make, converting the open-web scraping that trained a generation of models into a metered transaction. For the first time, the data commons that AI systems draw from has a tollbooth operator, and the operator sits in front of a large share of the web.

At the same time, the answer layer's reliability is failing in a genuinely new way. ProPublica invented a fake company to test Google's AI Overviews and watched the system cite the fabrication back as a real business - a fabrication-to-ratification loop distinct from ordinary hallucination, because the false information originates outside the model and gets laundered into authority by it. Tripadvisor's AI review summaries were found to soften and downplay traveler complaints, smoothing negative signal out of the record that people rely on to make decisions. Both cases show the same failure surface: a synthesis layer that presents confidence without preserving the friction that made the underlying information trustworthy.

The responses are arriving as fast as the failures. Wired reported on FLARE-AI, a harm-reporting framework built to catch exactly these downstream distortions before they compound. Noema published a serious argument for proof-of-personhood infrastructure - ways to establish that a participant online is human without surrendering identity - as the answer layer fills with synthetic text and synthetic reviewers. Libby, the library reading app, added an AI filter, and the Commonwealth writing prize faced an AI-authorship accusation, two small signs that human-verification fights are now reaching every institution that curates written work.

Taken together these are not separate messes. They are the early governance of a new commons being built in front of everyone, out of contested parts. The assumption cracking underneath all of it is that the open web was free to take and safe to trust by default - free because no one metered it, safe because a human wrote what you read. Both defaults are gone, and the tollbooths, the harm-reporting frameworks, and the personhood proofs are the first infrastructure of what replaces them. The friction is intense because the checkpoint is being installed while traffic is already moving through it, and every actor - the crawler-gatekeepers, the answer engines, the humans trying to prove they are human - is negotiating the terms at once. What emerges is a web where provenance and consent are priced in rather than assumed away, and that is a sturdier foundation than the one being lost.

The Other Side

For most of the industrial era, whoever wrote the rulebook for a new industry wrote themselves a permanent seat at the top of it. That is the bet underneath everything Washington is negotiating this week. OpenAI's pitch for a 5% federal stake, the voluntary standards deal built around a classified benchmark only the participating labs can see, the FTC warning companies that following state AI law could trigger federal enforcement - each move assumes a position written into the federal architecture now stays valuable later.

Other stories from the same week's cycle undercut the bet. Etched shipped a working inference chip against the ceiling one vendor was supposed to define. Open-weight models kept closing the capability gap, and several strong ones now run on ordinary hardware. A fence around frontier capability only holds if the capability stays scarce and stays inside the fence. Neither is true anymore.

The alternatives are already standing up in plain view. The EU and UK are building their own tech-sovereignty tracks rather than joining the compact, and sovereign compute is coming online outside it. The rulebook being drafted in Washington is one container among several, and the harder anyone tries to seal it, the more the rest of the world builds around it.

Imagine a founder in Lisbon or Nairobi in 2035, building a company on models that in 2026 were meant to sit behind a classified benchmark and a federal equity compact. She never applies to anyone for access. The capability is just there, running on hardware she owns outright, improving faster than any standards body could gate it. When she reads about the summer of 2026 - the pitched stake, the standards deal, the enforcement threats - it reads like a fight over the deed to a house while the ground it stood on was already sliding. The people inside that fight saw correctly that a settlement was being written. What they misjudged was whether it would hold. The hard year was the one when a handful of firms could still believe they were dividing the future among themselves. No better argument was needed to prove them wrong. The capability simply left the room, one open model and one new chip at a time, until the rulebook governed a scarcity that no longer existed.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a University of Minnesota team assembling SpudCell, a synthetic cell built from a 36-gene genome, and open-sourcing the recipe through a nonprofit instead of patenting it, Etched unveiling a working inference chip against Nvidia's assumed ceiling with a billion dollars in orders already booked, the FDA clearing the first regulatory T-cell therapy and extending CASGEVY gene therapy down to two-year-olds, Meta decoding full sentences from noninvasive brain recordings while Neuralink threads electrodes through the dura without opening it, the first instruments precise enough to watch a labor transition while it happens rather than reconstruct it afterward, and Firmus signing a twelve-year deal to bring its own gigawatt of clean power and pay its own way. There's also friction, and it's intense - finance and information sectors shedding roughly 28,000 jobs a month, Henrico County explaining a 25% rate increase to residents now hosting 37 data centers, Texas counting 32 proposed gas plants sited mostly where life expectancy already runs below average, OpenAI pitching a 5% federal stake beside a classified safety benchmark and an FTC warning that following state AI law could trigger enforcement, Google's AI Overviews citing a company that ProPublica invented back as real, Tripadvisor summaries softening the complaints travelers left, and Apple negotiating with blacklisted memory makers as a raid over smuggled Nvidia servers plays out in Taiwan. But friction generates sound, and sound is what makes a buried cost audible to the people already paying it. Step back for a moment and you can see it: the power bill, the safety standard, the labor effect, and the provenance of what a system tells you all moving from assumed externality to something counted, priced, and contested, the crown jewels of biology and silicon arriving as open protocols the moment anyone tries to lock them in a vault, and the checkpoint for the whole answer layer being installed while the traffic is already running through it. Every transformation has a breaking point. A load can crush the grid it bears down on... or, once it pays for what it draws, become the ground a sturdier one is built on.

AI Releases & Advancements

New today

  • Anthropic: Restored global access to Claude Fable 5 on July 1 following the lifting of US export controls, adding a new cybersecurity classifier that blocks the specific jailbreak technique identified by Amazon researchers in over 99% of cases; blocked requests are routed to Claude Opus 4.8. (Anthropic)
  • NVIDIA: Released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, an open-weight diffusion language model that splits the 30B Nemotron-3-Nano backbone into a frozen autoregressive context tower and a trained denoiser tower, achieving 2.42× generation throughput while retaining 98.7% of baseline benchmark quality. (Hugging Face)
  • Huawei: Open-sourced openPangu-2.0-Flash, a 92B-total / 6B-active MoE language model supporting reasoning, coding, and business automation, with weights, inference code, and training operators released on gitcode.com/ascend-tribe. (Pandaily)
  • xAI: Launched Voice Agent Builder, a no-code platform for creating Grok Voice-powered phone agents for customer support, sales, and scheduling; supports 25+ languages, 80+ voices, voice cloning from 2 minutes of audio, and costs $0.05/minute. (xAI on X)
  • Google: Launched Gemini Spark for macOS in beta for Google AI Ultra subscribers in the US, enabling the 24/7 agentic assistant to work with local files, sort and organize documents, automate desktop workflows, and connect to additional apps including Google Tasks and Google Keep. (Google Blog)
  • X (Twitter): Launched a hosted MCP server for the X API, enabling AI tools such as Claude, Cursor, and Grok Build to connect directly to X using a user's own account permissions without requiring developers to build and host their own MCP server. (X Developer Docs)
  • Google Research: Released TabFM, a zero-shot tabular foundation model trained on hundreds of millions of synthetic datasets that performs classification and regression on unseen tables in a single forward pass via in-context learning, now available on Hugging Face and GitHub. (Google Research Blog)
  • Base44 (Wix): Launched Base1, a proprietary LLM fine-tuned on Base44's own app-building session data and trained with reinforcement learning to generate visually distinct UIs; now available as a model choice alongside GPT-5.5 and Claude Opus 4.8 in the Base44 platform. (GlobeNewswire)
  • GitHub / Moonshot AI: Made Kimi K2.7 Code generally available to all GitHub Copilot users as a selectable model option. (GitHub Changelog)
  • Anthropic: Restored global access to Claude Fable 5 on July 1 following the lifting of US export controls, adding a new cybersecurity classifier that blocks the specific jailbreak technique identified by Amazon researchers in over 99% of cases; blocked requests are routed to Claude Opus 4.8. (Anthropic)
  • NVIDIA: Released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, an open-weight diffusion language model that splits the 30B Nemotron-3-Nano backbone into a frozen autoregressive context tower and a trained denoiser tower, achieving 2.42× generation throughput while retaining 98.7% of baseline benchmark quality. (Hugging Face)
  • Huawei: Open-sourced openPangu-2.0-Flash, a 92B-total / 6B-active MoE language model supporting reasoning, coding, and business automation, with weights, inference code, and training operators available on gitcode.com/ascend-tribe. (Pandaily)
  • xAI: Launched Voice Agent Builder, a no-code platform for creating Grok Voice-powered phone agents for customer support, sales, and scheduling; supports 25+ languages, 80+ voices, voice cloning from 2 minutes of audio, and costs $0.05/minute. (xAI on X)
  • Google: Launched Gemini Spark for macOS in beta for Google AI Ultra subscribers in the US, enabling the 24/7 agentic assistant to work with local files, automate desktop workflows, and connect to additional apps including Google Tasks and Google Keep. (Google Blog)
  • X (Twitter): Launched a hosted MCP server for the X API, enabling AI tools such as Claude, Cursor, and Grok Build to connect to X using a user's own account permissions without requiring developers to build and host their own MCP server. (X Developer Docs)
  • Google Research: Released TabFM, a zero-shot tabular foundation model that performs classification and regression on unseen tables in a single forward pass via in-context learning, trained on hundreds of millions of synthetic datasets; available on Hugging Face and GitHub. (Google Research Blog)
  • Base44 (Wix): Launched Base1, a proprietary LLM fine-tuned on Base44's own app-building session data and RL-trained to generate visually distinct UIs; now available as a model choice alongside GPT-5.5 and Claude Opus 4.8 inside the Base44 vibe-coding platform. (GlobeNewswire)
  • GitHub / Moonshot AI: Kimi K2.7 Code is now generally available to all GitHub Copilot users as a selectable model option. (GitHub Changelog)

Other recent releases

  • Anthropic: Released Claude Sonnet 5, now the default model on Free and Pro plans, delivering agentic performance close to Opus 4.8 - including multi-step browser use, terminal execution, and self-checking - at introductory API pricing of $2/M input and $10/M output tokens through August 31, 2026. (Anthropic)
  • Anthropic: Launched Claude Science in beta for Pro, Max, Team, and Enterprise users, a standalone AI research workbench integrating 60+ curated tools and connectors for genomics, proteomics, structural biology, and cheminformatics; coordinates specialist agents, submits HPC jobs, generates reproducible figures, and runs a reviewer agent that flags citation and calculation errors. (Anthropic)
  • Google DeepMind: Released Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), the fastest and lowest-cost model in the Nano Banana image generation family, delivering text-to-image outputs in ~4 seconds at $0.034 per 1,000 images; now live in Google AI Studio, Gemini API, and rolling out across Search AI Mode, Gemini app, NotebookLM, Google Photos, and Google Ads. (Google DeepMind)
  • Google DeepMind: Made Gemini Omni Flash (gemini-omni-flash-preview) available to developers via the Gemini API and Google AI Studio for the first time, enabling high-quality video generation and conversational editing from combined text, image, and video inputs. (Google DeepMind)
  • Mistral AI: Released Leanstral 1.5, a 119B-parameter MoE model with 6.5B active parameters and 256k context, fine-tuned for Lean 4 automated theorem proving and autoformalization; supersedes the March 2026 Leanstral and is available for free on the Mistral Labs tier. (Mistral AI)
  • Meituan: Open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model with ~48B active parameters, trained end-to-end on 50,000+ domestic Chinese AI accelerators across 35+ trillion tokens with 1M-token context; purpose-built for agentic coding and now available on GitHub and Hugging Face - previously available as the anonymous "Owl Alpha" stealth model that topped OpenRouter's developer charts. (LongCat Blog)
  • Cursor: Launched Cursor for iOS in public beta on all paid plans, bringing always-on cloud agents, remote control of desktop agents from the phone, voice input, diff review, and PR merging on mobile. (Cursor Blog)
  • Meta: Released Brain2Qwerty v2, a real-time non-invasive brain-to-text decoder achieving 61% average word accuracy (78% for top participant) from raw MEG brain signals without implants or surgery - compared to ~8% for prior non-invasive approaches - with training code for v1 and v2 released publicly alongside the v1 Nature paper. (MarkTechPost)
  • vLLM: Published Micro-Agent, a framework enabling multiple AI model API calls to collaborate and collectively outperform single frontier models on complex tasks, with a blog post and benchmark results released June 29. (vLLM Blog)
  • OKX: Launched OKX AI, an agent marketplace opening to developers on June 30 where AI agents can autonomously hire other agents, settle payments in stablecoins, and build portable on-chain reputations - following a closed beta with 50 early AI service providers. (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.

Read more