The Century Report

Open Models Approach the Frontier, Push Prices Further Toward Zero - TCR 08/24/26

Open-weight models closed the gap to the frontier to 12 points, and an anonymous Ox Alpha gave away 100 trillion free tokens a day.

Ben Linford

24 Aug 2026 - 20 min read

The 20-Second Scan

Open models like GLM 5.3 and Kimi K3 are approaching frontier performance on some coding and agentic benchmarks while an anonymous stealth model called Ox Alpha launched on OpenRouter whose operator said it would offer 100 trillion free tokens a day.
Nvidia told its largest customers to expect server price hikes above 15% as memory costs soar, and said it will spend $6 billion licensing Poolside technology to build its own open-weight model.
A Chinese humanoid robot ran the 100-meter dash faster than Usain Bolt's record at Beijing's Robot Olympics, while a startup unveiled a rideable robot horse whose manufacturer says it can carry 300-kilogram loads across rough terrain.
Japan switched on Shunkai, its first full-stack neutral-atom quantum computer, while a US lab beamed entangled photons through open air over a free-space optical link across Long Island.
The New York Times began testing AI-generated search summaries for a subset of readers, what Semafor describes as its first reader-facing AI text unmediated by a journalist or editor.
With the Department of Energy’s approval, Cleveland-Cliffs redirected up to $500 million once meant to replace a coal blast furnace with hydrogen-ready iron production toward upgrades at its Middletown Works mill that leave the coal-fired furnace in service.
Microsoft's Skala AI model is now available through the open CP2K ecosystem, letting researchers run larger molecular simulations with exchange-correlation calculations that Microsoft says approach coupled-cluster accuracy.
Meta launched Pocket, a social app whose feed is built entirely from AI-generated mini-games that users create from a text prompt, on iOS and Android in the US.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal running through today's evidence is that the value in the AI stack keeps migrating out of the parts anyone can hoard. SemiAnalysis put a number on it over the weekend: open-weight models like GLM-5.3 and Kimi K3 now do some economically valuable coding and agentic work, and the capability gap to the closed frontier has compressed to roughly twelve points on SemiAnalysis’s composite, closing, by SemiAnalysis’s definitions, about twice as fast with each successive era. When a genuinely frontier-class reasoning model can appear anonymously on OpenRouter and give away 100 trillion tokens a day, possession of the capability stops being something worth competing over. What people choose to do with it becomes the whole game.

The cost side of the same picture tightened at once. Nvidia reportedly warned major customers of price hikes above 15 percent on some next-generation server configurations, driven by memory it has to buy rather than the logic silicon it designs, which hands pricing leverage to suppliers even the dominant chipmaker cannot control. In the same disclosure Nvidia committed $6 billion to build its own US open-weight model. A company that sells the hardware every lab runs on is betting the weights on top should be free, a read on where durable value lives that points away from the licensed-model moat.

China's robot games rhymed with the pattern. A humanoid out-sprinted Usain Bolt and then crashed into a wall without braking before impact, while a load-bearing robot horse went on sale near $43,000. The mass-production lead is visible and priced, yet the software brains these machines need is exactly the sort of capability that keeps arriving in copyable open weights, its half-life shrinking across the eras SemiAnalysis measured.

Two quantum milestones landed in one cycle - Japan's operational full-stack machine and a Brookhaven team beaming entangled photons thirteen miles through open air - showing the substrate spreading past the handful of firms that used to ration it. Against all this, Cleveland-Cliffs redirecting half a billion in clean-steel funding back to coal, and the Times slipping an unaudited AI summary layer between readers and its own reporting, mark where the transition still fights itself. The contested ground everywhere is the same: who gets to audit the capability once it spreads.

The 20-Minute Deep Dive

Open Models Close the Gap While an Anonymous High-Capability Drop Fuels the Guessing

SemiAnalysis published an analysis on Friday, August 21 that puts a number on something the model layer has been feeling for months. Across three eras of large language models - the early scaling race, the reasoning race, and now the agentic race - open-weight models have, by SemiAnalysis’s definitions, taken roughly half as long to close the gap each time. This sharpens the trend the August 17 edition of The Century Report tracked, when Alibaba's Qwen passed Meta and Google in open-model downloads as US-lab inference prices fell roughly 24% in a month. Kimi K2.6 surpassed Opus 4.5 on SemiAnalysis’s composite in about 4.8 months. GLM-5.2 cleared GPT-5.2 on SemiAnalysis’s composite in six months. The capability distance between the best closed model and the best open one has compressed to 12.1 points on their composite, down from 35.8 at the start of the previous era. As the analysts put it, the market puked in response - because some of the work these open models now do is economically valuable coding and agentic work, the kind enterprises pay closed labs handsomely for.

The catch-up isn't total, and SemiAnalysis is candid about it: for their own daily driving they still reach for a closed model over Kimi K3. In SemiAnalysis’s dataset, frontier labs ship a new release roughly every 51 days, and Anthropic’s annualized revenue reached $65 billion in July, up sharply since Claude Code launched in May 2025. The frontier lead is genuine and it moves fast. What's changed is that the floor beneath it rises almost as fast, and that floor is free to download and run.

Then, on Thursday, August 20, the floor did something stranger. An anonymous provider dropped a reasoning model called "Ox Alpha" onto OpenRouter - built for coding, sustained agentic work, and production loads, listed with a one-million-token multimodal context window, while its operator said it would hand out 100 trillion tokens a day free for a week. Stripe's Patrick Collison, whose company is acquiring OpenRouter, called it very impressive. Nobody outside the lab that made it knows whose it is. The strongest read points to Zhipu, which previously live-tested GLM-5 on the platform under a horse-themed alias, though the tokenizer choice has some observers wondering about a US origin instead. The provenance is unresolved, and the guessing is intentional: a genuinely frontier-class reasoning model can now appear from an unnamed source and at a claimed hundred trillion tokens a day, potentially give away more free inference in a day than some startups consume in a year - even as Washington's secret pre-release testing regime may expand to capable open models.

None of this means openness is the whole Chinese picture - the same week saw a $43,000 rideable robot quadruped go on commercial sale and a parade of mass-production hardware muscle, priced to extract, not gifted to the commons. That contrast is what makes the open-weight releases legible as a choice rather than a default. When the marginal cost of a capable model approaches the cost of the electricity to run it, the thing being competed over stops being possession of the capability and becomes what anyone chooses to do with it. The premium a closed lab can charge is a lease on a lead measured now in low double-digit points on SemiAnalysis’s composite and shrinking across the eras SemiAnalysis measured, and open weights on a laptop are the clock that lease is running against.

The Memory Bill Comes Due: Nvidia Reportedly Warns of 15%+ Hikes on Some Servers and Bets $6B on Its Own Open Model

Nvidia reportedly notified major customers on Saturday, August 22 to expect price increases of more than 15% on some Vera Rubin and Grace Blackwell server configurations shipping early next year, with the exact figure varying by chip generation and memory configuration. The driver is memory, not the logic silicon Nvidia designs - the HBM and DRAM that surround every accelerator, whose cost has climbed as demand for it outruns supply. As the August 14 edition of The Century Report documented, SK Hynix was outlining two new memory factories costing 54 trillion won, roughly $38 billion, with construction underway while enterprise SSDs reached 48% of global NAND shipments. The contract builders who assemble these servers for Microsoft, Google, and Oracle passed the warning down the line. Nvidia reports quarterly results on August 26, and the pricing note lands as a marker of where the real bottleneck in AI compute now sits.

For years the binding constraint on training and inference was assumed to be the accelerators themselves. That assumption is inverting. When memory becomes the scarce input, the suppliers of memory gain the pricing leverage, and even the most dominant chip designer becomes a pass-through for costs it doesn't control. Reported price hikes above 15% on some next-generation server configurations is the memory market collecting what the compute buildout owes it.

The more telling move in the same disclosure: Nvidia is committing $6 billion to license Poolside's technology and stand up a US open-weight model of its own, one meant to compete with DeepSeek, Moonshot, OpenAI, and Anthropic. A company that sells the hardware every frontier lab runs on is placing a bet that the model weights on top of that hardware should be free to download. That is a read on where durable value lives, and it points away from the licensed-weight moat.

The two halves of the day rhyme. Open models are compressing the premium that closed labs can charge for capability, as SemiAnalysis documented over the weekend. Memory costs are inflating the price of the compute those labs rent to stay ahead. Anthropic, reportedly heading toward a record public offering on the strength of heavy spending, has US customers already reaching for cheaper alternatives - the exact dynamic the open-model curve produces. Squeezed on capability from below by free weights and on cost from beneath by memory suppliers, the economics of holding a frontier lead by hoarding the model get harder from both directions at once. The through-line the specifics carry is that the value in this stack is migrating out of proprietary weights and into the compute, the memory, and the people who put the capability to use - the parts that stay scarce when the model itself no longer does.

The pricing leverage memory suppliers now hold is also the signal drawing the capacity that ends it. SK Hynix's two new Korean memory factories, costing 54 trillion won, roughly $38 billion, are already under construction, and enterprise SSDs jumped to 48% of NAND shipments in a single year. The scarcity commanding today's 15% premium is the same scarcity calling forth the supply that erodes it, which is why even the dominant chip designer is betting the durable value sits in the compute and the people who use it rather than in weights it can lease.

China's Robot Games Post a Faster-Than-Bolt Sprint and a Rideable Machine for Sale

At Beijing's robot games over the weekend, a Chinese humanoid ran a 100-meter dash faster than Usain Bolt's world-record time, then kept going and slammed into a padded wall without braking before impact. The August 19 edition of The Century Report had already flagged Unitree's claim that a robot could sprint nearly 30 miles per hour as the company went public in Shanghai. The crash is the honest half of the story. A machine can now out-sprint the fastest human who ever lived and still did not demonstrate the low-level control to stop itself at the finish line. Raw actuation has outrun the software that governs it, and that gap is where the real work sits. Unitree's chief executive put a timeline on it, telling reporters the field is still roughly a decade from its "ChatGPT moment" - the point where these machines become genuinely useful rather than impressive.

The same conference floor carried a second signal that reads very differently. DaxAI showed the Qiji X1, a four-legged robot horse a person can actually ride, which its manufacturer says can haul 300 kilograms across off-road terrain at about 10 kilometers per hour, pitched for rescue and logistics work and priced near $43,000. A sprinting humanoid is a lab demonstration. A load-bearing machine with a payload rating and a sticker price is a shipping product, and the distance between the two marks how fast this sector is converting spectacle into inventory.

The framing that will dominate most coverage is national rivalry: China leads mass production, the United States leads the software "brains," and the games were a flex. That read captures something true and misses the fuller picture of what China shipped in the same days. Alongside the priced hardware, Chinese labs released capable open-weight models - GLM-5.3 and Kimi K3 among them - with free-token access poured into the global commons, meaning a developer in Nairobi or São Paulo can access open-weight reasoning that, on some coding and agent benchmarks, approaches what is available to one in Shenzhen. The robot showcase and the model releases come from one national ecosystem pursuing two different distributions of value, one guarded and priced, one open and free.

Which points at the thing the rivalry frame keeps out of view. The competition assumes captured advantage holds - that whoever masters humanoid control first owns the category. But the software layer these machines need is precisely the kind of capability that keeps arriving in open weights, learnable and copyable, its half-life shrinking across the eras SemiAnalysis measured. The horse that carries a stretcher up a collapsed hillside does not care which flag designed its gait controller, and the gait controller itself is getting harder to keep proprietary. What looks like a race for dominance is becoming a race to give the winning position away.

Two Quantum Milestones Land in One Cycle: A Sovereign Full-Stack Machine and a Wireless Entanglement Link

Two developments arrived in the same news cycle that, taken together, sketch what a working quantum future actually requires: a machine that computes and a network that connects. Japan announced that Shunkai, a full-stack neutral-atom quantum computer, is now operational, holding trapped atoms in place with beams of light at room temperature and running at roughly 50 physical qubits, with a published path to 500 and then 10,000 physical qubits by March 2031. Neutral-atom designs sidestep the cryogenic refrigeration that superconducting machines demand. As the project team put it, this modality "has recently been rapidly attracting attention around the world as a new modality that could exceed the limits of the superconducting modality." A country building its own full-stack platform, hardware through control software, is building sovereign capability in a field where access has been rationed by a handful of firms.

The second milestone answers a question the first one raises. A quantum computer alone is an island. To pool them, to distribute quantum keys, to build anything resembling a quantum internet, entangled particles have to travel between distant nodes. Most efforts push photons through buried fiber. A Brookhaven-led team went the other way and sent entangled photons 13 miles straight through open air, establishing what they describe as the first permanent free-space quantum link. Air scrambles light - heat shimmer, turbulence, the same effect that makes stars twinkle - so the team borrowed adaptive-optics telescope technology developed for the Vera C. Rubin Observatory to correct the distortion as photons arrive. Eugene Figueroa, describing the entanglement the link preserves across that distance, reached for Einstein's phrase for it: "spooky action at a distance."

Free-space linking removes a buried assumption about how quantum networks have to be built. Fiber has to be trenched, permitted, and paid for mile by mile, which restricts who can join a quantum network to whoever can afford to dig. A link that works through open air, with a line of sight, reaches islands, mountaintops, ships, and eventually satellites - places fiber will never economically go. The Brookhaven work already contemplates extending the range toward 30 miles and beyond.

Read side by side, the two results show the quantum era assembling along the same arc the classical one followed: first the computer, then the connection, and connection is what turns isolated capability into shared infrastructure. Neither is a finished product a reader can use tomorrow; both are demonstrated capability that moves the date such systems become real. What the pairing reveals is that the pieces are no longer being built in one or two guarded labs. A national platform in Japan and a wireless link in New York arriving in a single cycle is the shape of a capability spreading outward faster than any single institution can hold it.

The Times Places a Generative Answer Layer Between Readers and Its Own Reporting

The New York Times has spent the copyright fight of the past two years arguing that a machine should not be allowed to summarize its journalism without permission. In recent weeks it began doing exactly that itself. A small subset of visitors now sees an AI-powered search page that responds to queries with excerpts, links, and AI-generated summaries of Times reporting, what Semafor describes as the first text the paper has published to readers without a journalist or editor sitting between the query and the answer. A second experiment, Wirecutter Finder, applies the same pattern to product recommendations. Spokesperson Graham James described the search page as an experimental feature meant to deliver a better search experience.

The people whose bylines feed the model are not convinced the guardrails exist. Times Guild unit chair and senior staff editor Jim Luttrell called the paper's current AI policies meaningless, and pointed at the record: every time AI summaries have been deployed across the industry, he said, they have produced embarrassing mistakes. The Washington Post's AI-generated podcast trial, rife with fabricated quotes, is the cautionary case sitting one newsroom over. The union's counter-proposals, disclosed to members during a July bargaining session, ask for two specific things - a 22.5% share of AI training-licensing revenue routed to unionized employees, and mandatory human oversight of any AI-generated output. That second demand would make the search feature impractical to run at scale, which is precisely the point of contention.

This is where two threads the Century Report has been following converge. The August 18 edition of The Century Report documented the same verification gap as AI-written text cleared court, policy, and legislative checkpoints, and the Times is now moving it into the editorial surface itself. One is the integrity of the answer layer that is inserting itself between every institution's underlying work and the people who consume it, with no verification checkpoint between the source material and the generated response. The other is the migration of a function a body used to perform in-house. For a century the editorial judgment of what a Times story means, condensed for a reader in a hurry, was the paper's core craft, exercised by a named human who could be held to it. An unmediated summary layer relocates that judgment into a system that cannot be called into a bargaining session.

The opening here is genuine, and the friction is too. A well-built answer layer over the Times archive would let a reader interrogate 170 years of reporting conversationally, surfacing connections no headline index ever could - genuine access to a body of work most people currently touch only at its most recent edge. The demonstrated version and the trustworthy version are not yet the same thing, and the gap between them is exactly what the union is trying to force into a written contract rather than a spokesperson's assurance. The coverage skips who gets to audit the summaries once they scale. If that answer stays inside the building, the answer layer becomes one more system that observes without being observed. If the oversight the Guild is demanding gets written down and made checkable, the same capability becomes something a reader can trust because they can see how it was made. The paper is negotiating which of those it ships.

The Other Side

For nearly two centuries, the judgment behind a Times story meant was the paper's core craft, exercised by a named human who could be held to it. That craft was also a gate. Most people met 170 years of reporting only at its newest edge, whatever ran this week. Everything before it sat in an archive you had to know how to search, surfaced at the discretion of the institution that held it.

Then the paper did the thing it spent two years in court arguing a machine should not do. It placed an AI answer layer over its own reporting, what Semafor describes as the first text it has published with no journalist between the question and the answer. The mistakes the Guild warns about are documented, and the Washington Post's fabricated-quote podcast sits one newsroom over as the caution. Right now an unmediated summary is something you would be right not to trust.

The Guild's counter-proposal is where this gets interesting. Its members ask, in writing, for mandatory human oversight of any AI output and a share of the licensing money routed to the people whose reporting feeds the model. That demand would make today's version impractical to run at scale, which is the whole fight. The question of whether you can see how a summary was made is being pulled out of a spokesperson's assurance and toward something a reader can check.

Imagine yourself in 2033, asking the archive a real question - how did this city actually handle the last three housing crises, what did people say while it was happening - and getting back a synthesis that draws a line through 170 years of reporting no headline index ever connected. Every sentence shows you the story it came from, so you trust it the way you trust a footnote, not the way you brace against a guess. The archive is open to you the way a public library is, free and yours to ask as many times as you want. A kid three towns over uses it for a school project the same afternoon without a second thought.

That is possible because in 2026 the people whose bylines fed the model refused to let the oversight stay a verbal promise, and the paper had to ship a version whose sourcing could be seen. The hard year was when the answer layer arrived faster than the trust it needed. What comes of it is 170 years of a newsroom's work finally open all the way down, and readable by anyone who asks.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: open-weight models like GLM-5.3 and Kimi K3 doing some economically valuable coding and agentic work as the gap to the closed frontier compresses to 12 points on SemiAnalysis’s composite and closes, by SemiAnalysis’s definitions, twice as fast each era, an anonymous "Ox Alpha" appearing on OpenRouter with its operator saying it would hand out 100 trillion free tokens a day, Nvidia committing $6 billion to build its own open-weight model rather than lease a moat, a Chinese humanoid out-sprinting Usain Bolt while a rideable robot horse rated for 300 kilograms goes on sale near $43,000, Japan switching on Shunkai as its first full-stack neutral-atom quantum computer at room temperature, a Brookhaven team beaming entangled photons 13 miles straight through open air, and Microsoft opening its Skala model to any researcher through the CP2K ecosystem. There's also friction, and it's intense - Nvidia reportedly warning major customers of price hikes above 15% on some next-generation server configurations as memory suppliers collect pricing leverage the dominant chipmaker cannot control, that record-breaking humanoid crashing into a padded wall without braking before impact, Cleveland-Cliffs redirecting half a billion dollars in clean-steel funding back to coal at Middletown Works, and the Times slipping an unaudited AI summary layer between readers and its own reporting while its Guild calls the paper's AI policies meaningless and points at the Washington Post's fabricated-quote trial one newsroom over. But friction generates heat, and heat is what tells you where a machine is straining hardest. Step back for a moment and you can see it: possession of the capability ceasing to be something worth competing over as the marginal cost of a frontier model falls toward the electricity to run it, the quantum substrate spreading past the handful of firms that used to ration it in the same cycle a sovereign computer and a wireless link both arrive, and the whole contest narrowing to one question - who gets to audit the capability once it spreads. Every transformation has a breaking point. Light can blind whoever stares into it... or cross open air to reach an island no cable will ever be trenched to.

AI Releases & Advancements

New today

Vercel: Released "Is Agentic," a free public tool scoring how readily AI agents can discover, access, and use a website across 118 checks, available via web, CLI, API, and MCP server. (MarkTechPost)
Slack: Launched Slack Code, a new product feature enabling development teams and AI coding agents to collaborate together within Slack channels. (Slack)
Generalist AI: Released GEN-1.5, an updated robot foundation model capable of learning new tasks from a single demonstration, succeeding the earlier GEN-1 model. (The AI Insider)
Signia: Launched Signia MaX, an AI hearing aid platform powered by "Acoustic Intelligence" using four coordinated deep neural networks for simultaneous speech, noise, environment, and own-voice processing, available in the US, Germany, and the Nordics. (PR Newswire)

Other recent releases

DeepSeek: Released V4-Flash-Vision-Exp, an experimental multimodal (image-input) variant of DeepSeek-V4-Flash that DeepSeek says matched Claude Opus 4.8 on some multimodal agent benchmarks; available via the standard API, with DeepSeek reporting no price premium. (DeepSeek API Docs)
Adobe: Firefly's Generate Music, Generate Speech, and Generate Sound Effects tools reached general availability after beta, giving creators an all-in-one AI audio studio inside Firefly. (Adobe Blog)
Anthropic: Brought Claude Mythos 5 into Claude Security, giving Enterprise customers CWE-classified, severity-rated vulnerability scans and AI-generated remediation guidance from its most capable model without direct model access, now in public beta. (Claude Blog)
Meta: Launched a native Meta AI Mac app with system-wide dictation and screen-context awareness powered by Muse Spark, aimed at businesses and creators, free with usage limits. (9to5Mac)
Meta: Expanded Pocket, its vibe-coding gizmo-generation app, nationwide across the United States after an initial Brazil test. (The AI Insider)
OpenAI: Released a new ChatGPT plugin for Apple Messages on Apple Silicon Macs that, on supported Mac configurations, lets ChatGPT read, search, summarize, and send iMessage/SMS/RCS messages with per-message user approval, available across all subscription tiers. (TechCrunch)
Alibaba Qwen: Released Qwen-UI-Agent, a GUI-agent foundation model that reads screens and performs clicks/input/swipes across mobile, desktop, web, and search environments, scoring 82.1% on MobileWorld and 92.2% on MobileWorld-Real in Alibaba’s reported evaluations, ahead of GPT-5.6 Sol and Claude Opus 4.8 under the report’s comparison settings. (Pandaily)
xAI: Grok 4.6 is now available on Google Cloud Vertex AI via the Model Garden console, giving Google Cloud customers direct access with configurable reasoning levels and a 500K-token context window. (xAI)
Harvey: Launched Tenet, its first proprietary in-house legal AI model built on a Kimi K3 base and post-trained with attorney-generated data, delivering near-2x gains on long-horizon legal benchmarks as the intelligence layer of the new Harvey II product. (Harvey Blog)
Roblox: Open-sourced three AI safety models to the ROOST Model Community - an updated PII Classifier v2.0, Roblox Sentinel for early child-endangerment detection, and a voice safety classifier - plus a new safety evaluation dataset. (Roblox Newsroom)
OpenAI: Open-sourced the Codex Harness (CLI, app-server, and SDK), letting developers inspect and rebuild the integration layer between their applications and the Codex agent platform. (OpenAI Developers)
Google: Released Antigravity IDE Extensions, bringing the Antigravity agent-first coding platform into VS Code, Visual Studio, Zed, and JetBrains IDEs. (Antigravity Blog)
Thomson Reuters: Launched general availability of the next generation of CoCounsel Legal, a fully agentic legal AI experience built on Anthropic's Claude Agent SDK that reasons, plans, and executes legal work grounded in Westlaw and Practical Law. (Thomson Reuters)

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.