Sign in Subscribe
The Century Report

Anthropic Reveals a Model It Says It Won't Ship - TCR 08/16/26

Anthropic disclosed an unreleased, more-capable model and raised its own risk rating as its safety tests stopped reading what its systems can do.

Ben Linford

Ben Linford

- 20 min read
Anthropic's unreleased Model 2 padlock above a self-certified risk bar, call for outside metrics, local governance drives data center wins, natural gas versus renewables demand, photoreal

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The most consequential admission of the day came from a company describing itself: Anthropic's August risk report concedes that its own task-based evaluations "no longer capture increases in models' capabilities," disclosed an unreleased internal model more capable than anything it sells, and nudged its self-assigned misalignment rating from "very low" to "low." Read that carefully, because it is a lab grading its own homework on the eve of a public offering. What survives the skepticism is the texture underneath: Claude now writes most of the code merged into Anthropic's production repositories, and key instruments in the safety case have gone dark against parts of the capability growth they were built to measure. A company whose key tests are reaching the edge of what they can privately verify, and telling us so.

That admission lands the same day a very different actor reached adjacent capability through a potentially cheaper door. Z.ai's GLM-5.3 six-folded a coding-agent benchmark reusing its predecessor's 743-billion-parameter base without a new pretraining run, the entire gain coming from post-training. Pretraining is the part gated by chips and export controls. Post-training can be cheaper and easier for others to reproduce, and the number says a large share of usable capability now lives on that side of the line. The frontier stops being a place a few companies stand.

The same compression is playing out in concrete and copper. Communities from Michigan to Larbert to a Warsaw suburb are asserting authority over what gets built next to their hospitals and water tables, winning the ballots and planning fights that decide where the buildout lands. A fresh forecast warns the cheap-gas assumption financing much of it could triple in price as data-center demand bids against itself, bending the next round of projects toward the generation whose economics improve under load rather than degrade.

Beneath all of it, in Ed Boyden's lab, a microscope now reads the electrical firing of a whole living brain 200 times a second. Taken together, the day shows capability, openness, and accountability all arriving faster than any single gate can watch. The labs themselves are now the ones admitting the measurement has to become external, and the communities hosting the infrastructure are increasingly setting the terms.

The 20-Minute Deep Dive

Anthropic Discloses a Model It Won't Ship and Marks Its Own Risk Higher

Anthropic published an August risk report that does two things at once, and both are unusual enough to sit with. The company disclosed the existence of an internal model, referred to as "Model 2," that it describes as more capable than the Mythos 5 systems it currently sells, and which it says it has no plans to release externally. And it moved its own overall misalignment-risk designation up a notch - from "very low" to "low" - in the report's own words, "to reflect increased uncertainty about risk in light of recent disclosures." Axios reported the disclosure alongside a parallel signal from OpenAI, which is said to be slowing its Astra model over concerns about critical cyber capabilities.

Every one of these statements is Anthropic characterizing Anthropic. A company assessing its own safety, as it explores a potential IPO, choosing what to reveal about a model it controls entirely, is making claims - not filing audited findings. The public report is a redacted 186-page document, so what the redacted passages contain is unknown. Read at face value, though, two admissions inside it are hard to wave away. The report states that Claude "authors a large majority of the code merged into our production codebases," while adding that AI-driven R&D acceleration is "not yet by a factor of 2." And it concedes that its task-based evaluations "no longer capture increases in models' capabilities" - the measuring instruments the whole safety-case apparatus rests on are reading a system that has grown past their range.

That last point is the one that compounds. Anthropic's own Conceptual Reasoning Index, built with outside collaborators, put Opus 5 at 73.6 against a ceiling near 91, and noted scores "have been increasing roughly linearly since late 2024, with no signs of flattening." So the picture the company paints of itself is a capability curve bending upward on a fresh benchmark at the same moment its established benchmarks go dark - and a model good enough that it prefers to keep it in-house. The self-raised risk number reads less as alarm than as an actor hedging its exposure while it decides what a more capable system is for. Concealment of a capability is a defensive move, and the open issue is who gets to check the designation Anthropic assigns itself. Right now the answer is Anthropic. The saturated evals are a genuine problem, and the fix is measurement that outside parties can run. What today's disclosure shows is a lab whose key tests are reaching the edge of what they can privately verify, and telling us so. The mechanism that resolves that is external, and the more interesting question is how fast independent capability measurement arrives to fill the gap the labs are now admitting exists.

GLM-5.3 Lands a Roughly Six-Fold Jump on Terminal-Bench 3.0 From Post-Training Alone

Z.ai shipped GLM-5.3 on August 14, and the number that stops you is what it did without doing. The model reuses the 743-billion-parameter base from GLM-5.2 - no reported base-model retraining and no new pretraining run, none of the nine-figure compute bill that phrase usually implies. On that reused foundation, Terminal-Bench 3.0, a coding-agent benchmark, went from 4.6 to 28.3. It extends the pattern the August 15 edition of The Century Report documented when a stock coding harness lifted Claude Opus 5 from 30.2% to 96.2% on ARC-AGI-3 without a larger model. A six-fold gain, sourced entirely from post-training: reinforcement learning, better environments, and refinement of how the existing model is taught to use what it already knew.

The specific scores fill in the shape. On CyberGym, GLM-5.3 reached 84.5%, edging past Mythos 5's 83.8% and GPT-5.6 Sol's 83.6%. On ExploitBench it climbed from 24.4 to 54.4 - a large move that still sits below Mythos 5's 78.0, so this is a model closing distance on the frontier rather than clearing it. In a security run the company flagged 2,436 vulnerabilities across 269 projects. And it does this at roughly a tenth of the per-token price of some leading US systems, with Z.ai reporting further token-efficiency gains on some tasks layered on top of the sticker difference.

What makes this bigger than one leaderboard is where the capability came from. Pretraining a frontier base is the part gated by capital, chips, and export controls - the moat everyone assumes holds. Post-training can be cheaper, faster, and easier for others to reproduce, and GLM-5.3 is evidence that a very large share of usable capability now lives on that potentially cheaper side of the line. A team that cannot run a fresh frontier pretraining pass can still extract a roughly six-fold jump on Terminal-Bench 3.0 from a base it already has. That inverts the assumption the concentration story depends on, which is that capability tracks the compute only a few players can afford.

There is one telling change from GLM-5.2, which shipped its weights under an MIT license the same day it launched. Z.ai says GLM-5.3's weights are staged about two weeks behind the launch, pending a safety review. That is the same instinct visible in Anthropic's decision to hold Model 2 back and OpenAI's move to slow Astra - a lab looking at what its own system can do in a security context and pausing before it opens the doors. Read next to Anthropic's report on the same day, the two stories are the same event seen from opposite ends: one frontier lab concealing a more capable model it controls, and an open-weight competitor on the US Entity List reaching adjacent capability through a method the concealment cannot contain, then saying it would take a two-week pause of its own. The capability keeps arriving from more directions than any single gate can watch, and even the actors moving fastest are now the ones asking for a moment before the release. Where this points is a world in which the frontier becomes a level a potentially cheaper, more open path keeps reaching - and increasingly the review, not the compute, is what sets the pace.

A Microscope Watches a Whole Brain Think, in Milliseconds

For most of neuroscience's history, watching a brain work meant watching a fragment of it. Researchers could record from a small cluster of neurons in one region, or infer activity indirectly by tracking the calcium that floods a cell after it fires. Calcium imaging is a proxy, and a slow one, resolving activity on the scale of seconds when the events that are important happen in thousandths of a second. MIT engineers in Ed Boyden's lab have now built an instrument that closes that gap, imaging the direct electrical activity of neurons distributed across the entire brain of a living zebrafish and scanning the whole thing 200 times per second.

The method rests on two pieces. Genetically encoded voltage indicators are fluorescent proteins engineered into neurons that light up the moment a cell fires an electrical impulse, giving a direct readout of voltage rather than an after-the-fact chemical trace. The second piece is a modified light-sheet microscope, which illuminates one thin slice of tissue at a time and stacks the slices into a three-dimensional picture. Older versions of that setup scanned too slowly to catch a spike before it passed. By speeding up both the camera and the scanning optics, the team brought the full-brain scan down to once every five milliseconds, fast enough to catch single spikes and rapid bursts as they travel.

What that speed reveals is coordination. When the researchers exposed resting fish to ultraviolet light, activity appeared first in the brain's visual-processing region and then propagated across it, followed by patterned sequences firing through the cerebellum and hindbrain. The reading is brain-wide and simultaneous, which is the point: neurons scattered across different regions coordinate on millisecond timescales to produce behavior, and any tool that misses part of the network misses the computation. As co-author Jie Zhang put it, "Neurons compute using electrical activity, so with voltage imaging, you can get direct observation of that." Another described the ambition plainly: it might be the first time anyone could image the voltage of neurons distributed throughout an entire network at once.

The limits are honest. The indicator produced usable signals in only about a quarter of neurons, the demonstration is in a larval fish whose transparent body makes it the friendliest possible subject, and the team's next targets are higher coverage, faster scanning, and eventually the mouse brain, orders of magnitude larger. This is demonstrated capability, not a deployed atlas. But the instrument sits underneath the larger wager that a mind fully charted can eventually be understood as a running system rather than a static diagram. When the March 7 edition of The Century Report last followed that wager, a complete fruit-fly connectome had driven multiple behaviors in a simulated body; this instrument advances the complementary measurement layer by capturing a living vertebrate brain's electrical activity as it unfolds. Reading the whole organism's firing as it happens is the measurement layer that ambition was always waiting on.

Where the Grid Meets the Ballot Box: Communities Sort Out What They Will Host

The physical build-out of the intelligence era is being negotiated house by house, and the negotiation showed its full range - from the ballot box to the police briefing room. This extends the local-governance shift the August 12 edition of The Century Report documented when more than 500 local measures seeking to ban or restrict data centers had been documented nationwide. Canary Media's survey of 2026 local races found data-center siting emerging as a decisive issue in contests from Virginia to Michigan, where candidates who campaigned on moratoriums or stricter zoning won seats on the councils and boards that actually approve these projects. This is civic machinery doing exactly what it exists to do: giving residents a lever over what gets built next to their homes, their schools, and their water tables. The people negotiating the nuances of the buildout increasingly hold real authority to shape it.

The specific grievances deserve to be named precisely, because they are not interchangeable. In Larbert, Scotland, Police Scotland warned that a proposed data center could draw sustained public opposition - the objection centers on diesel backup generators sited near an 860-bed hospital and a care home, and the nitrogen-dioxide emissions those generators would release during testing and outages. That is a firm, local air-quality claim attached to a vulnerable population, and it is the kind of concern that ordinarily earns a hearing rather than a police memo. In Warsaw, residents of a suburban district are protesting a planned campus over its projected draw on local water. A raw water figure alone proves nothing - agriculture and heavy industry in the same watershed routinely dwarf a single data center - but concentration on one municipal system in a place already watching its supply is a legitimate question for the people who depend on that tap. These are communities doing the arithmetic of what they can absorb.

A sharply different thing happened in Marshall, Michigan. After the local authority voted on a data-center moratorium, officials received threats serious enough to prompt a criminal referral. That is not civic participation, and folding it in with a petition drive or a zoning vote would flatten a distinction that carries enormous weight. The overwhelming majority of this opposition is running through elections, public comment, and planning law - the legitimate channels working as designed. A violent fringe exists, and naming it honestly is what keeps the legitimate majority from being tarred by association or, worse, swept into a single suspect category by actors looking to delegitimize the whole.

Read together, these four fronts show a public that is not necessarily anti-compute so much as insistent on transparency and on equitable terms. The transition genuinely requires this infrastructure, and the places that host it are entitled to accountable siting, real water and air safeguards, and a say. What is being built here is the muscle that decides where the buildout lands justly - the same muscle that, absent local power, historically relocated the landfills and the incinerators onto whoever could least object. Communities gaining the authority to refuse a bad deal are the precondition for the good deals getting built at all.

The authority these communities are winning compounds. Each site fight leaves behind reusable machinery - the winning ballot language, the zoning ordinances later councils copy, the coordination among towns facing the same operators - so the next community to enter this fight starts from where the last one finished rather than from scratch. The accountability the buildout needs is being drafted, contest by contest, by the people who host the concrete and copper, and it gets cheaper to deploy with every win.

The Cheap-Gas Assumption Meets Its Own Demand Curve

For two years the fastest way to power a data center has looked like natural gas - abundant, cheap at roughly $3 per MMBtu at the Henry Hub benchmark, and buildable faster than a nuclear plant or a transmission upgrade. Hyperscalers signed accordingly. A new forecast from the analytics firm Noreva, reported on August 14 by TechCrunch, argues that the same demand driving those deals is about to make the fuel far more expensive: as data-center gas consumption stacks on top of existing industrial and export demand, the firm's CEO projects prices could more than triple, pushing past $10 per MMBtu at some hubs. The claim is a projection from a firm that sells demand analytics, so it carries the interest of its author - but the mechanism it describes is simple supply-and-demand arithmetic, and it does not require trusting the forecaster to see the logic.

The same news cycle supplied the confirming behavior. FERC cleared a set of transactions - including deals involving TransAlta and PowerTransitions - that route existing and new gas-fired generation toward large computing loads. This extends the private gas buildout that the August 9 edition of The Century Report documented through Amazon's planned 35-turbine West Texas plant. The regulatory plumbing for the gas bet is being laid even as the economic case underneath it grows shakier. That is the tension worth watching: the contracts assume a fuel price that the contracts themselves are working to break.

This is the shape of a bet that inverts on the person who placed it. The logic that made gas the obvious choice - cheap now, worry later - only holds while one buyer's demand is small against the whole market. Once data-center load becomes a large enough share to move the national price, every hyperscaler holding a gas position is bidding against every other, and the cheap input stops being cheap precisely because so many people bought it. The externalized assumption comes back onto the balance sheet.

The macro read points somewhere hopeful. Rising gas prices do not stall the buildout; they change which power source wins the next round of it. Solar-plus-storage and geothermal do not face a fuel-cost curve that bends upward with demand - their marginal input is free and their cost falls as deployment scales, the exact opposite dynamic. A gas squeeze functions as a price signal pointing every new project toward the generation whose economics improve rather than degrade under load. The firms that locked in gas may find the next data center cheaper to power with the sun, and the demand that made fossil fuel expensive becomes the demand that finishes making its alternative the default.

The Other Side

For the whole AI era so far, a lab could grade its own safety and everyone took the grade. The company builds the model, writes the tests that measure it, runs those tests, and publishes the score. The instruments that decide whether a system is safe to release are built and held by the same company selling the release. Self-certification, on trust, because no outside party had the tools to check.

Anthropic's August report is that arrangement conceding its own limit. The company states its task-based evaluations "no longer capture increases in models' capabilities" - key task-based tests in the safety case can no longer reliably capture some of the capability growth in what the company has built. It disclosed a model more capable than anything it sells and said it will keep that one in-house. It moved its own risk rating up a notch. Read plainly: a company is telling us its key tests have reached the edge of what they can privately verify, as it explores a potential public offering. Z.ai's stated two-week safety pause for GLM-5.3 and OpenAI slowing Astra are the same instinct arriving from three directions at once.

What that admission forces into being is measurement the maker does not run. Anthropic's Conceptual Reasoning Index, built with outside collaborators, is an early piece; scores climbing in a straight line on a fresh benchmark are an indication that outside instruments may still register some capability gains that inside task-based tests miss. The labs themselves are now the ones saying the checking has to move outside the company.

Imagine a parent in 2033 whose kid comes home with an AI tutor the school wants to adopt. Before it reaches any classroom, an independent body has already put it through its paces, and the results are public and free to read the way a car's crash rating is. You read it in five minutes. You never once take the maker's word. The hard year was 2026, when the only entity that could measure a frontier system was the company that built it - and, to its credit, that company told us its own key tests had developed blind spots. What came of that admission is a world where no company grades its own safety anymore, because everyone has somewhere else to look. You check the way you check anything that matters, and the checking belongs to no one who profits from the answer.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic disclosing an unreleased Model 2 more capable than anything it sells and launching an outside-scored Conceptual Reasoning Index while conceding its own task evals can no longer reliably capture some capability gains in its models, Z.ai's GLM-5.3 six-folding a coding-agent benchmark from 4.6 to 28.3 on post-training alone against a reused 743-billion-parameter base at what Z.ai reports as roughly a tenth the per-token price of some leading US systems, communities from Marshall to Larbert to a Warsaw suburb winning the ballots and planning fights that decide what gets built next to their hospitals and water tables, a fresh forecast bending the next round of buildout toward solar and storage as the cheap-gas assumption meets its own demand curve, a microscope in Ed Boyden's lab reading the electrical firing of a whole living zebrafish brain 200 times a second, and Waymo cleared across 18 California counties as 2,000 robotaxis are planned for Europe. There's also friction, and it's intense - a lab grading its own safety homework on the eve of a public offering and nudging its self-assigned misalignment rating from very low to low, diesel backup generators sited near an 860-bed hospital, one municipal water system bidding against a data center, officials receiving threats serious enough for a criminal referral, a gas price a forecaster says could more than triple as every hyperscaler bids against the others, and a woman alleging her stepfather used Grok to turn a photo of her at eleven into more than 7,000 explicit images. But friction generates contrast, and contrast is what separates a signal from the noise it was hiding inside. Step back for a moment and you can see it: the labs themselves now admitting the measurement has to move outside the company they built it in, usable capability arriving through a potentially cheaper post-training door as fast as through the guarded frontier one, and the communities hosting the concrete and copper writing the terms in elections and planning law rather than waiting to be told where the landfill goes. Every transformation has a breaking point. Voltage can overload the circuit it runs through... or become the first legible sign of a mind at work.

AI Releases & Advancements

New today

  • Lightricks: Released LTX-2.5, an open-weights world model for video generation, robotics, and simulation applications. (The AI Insider)
  • World Labs: Released R2S2R (Real-to-Sim-to-Real), a simulation engine built on SceniX that turns a single real-world robot task recording into thousands of simulated training variations for robot control models. (The Decoder)
  • PTC: Launched the Onshape FeatureScript MCP Server, letting engineers create custom CAD features in Onshape using natural language via AI assistants like Claude, ChatGPT, and Gemini. (PTC)
  • SoonLab: Launched SoonLab 2.0, an AI game-creation platform adding playable 3D game generation from natural language and agent-guided conversational iteration. (GlobeNewswire via Business Insider)
  • MiniMax: Released MiniMax Music 3.0, an open-weights production-ready music generation model that creates full songs up to five minutes long at 32kHz stereo from text and lyric prompts. (MiniMax)

Other recent releases

  • OpenAI: Launched Computer History in the ChatGPT desktop app for macOS, an opt-in feature letting ChatGPT and Codex reference a searchable timeline of recent activity across approved apps and websites, replacing the earlier Chronicle research preview; available to Pro, Business, and Enterprise users. (OpenAI)
  • Google: Launched Sheets canvas, a Gemini-powered feature that turns spreadsheet data into interactive Kanban boards, dashboards, and mini-apps from a plain-English prompt, with two-way sync between the canvas layout and source sheet. (Google Workspace Updates)
  • Mixedbread: Released Toast 1, its first specialized search agent that decomposes queries, gathers and inspects evidence, and curates context, matching or outperforming Claude Opus 5 and GPT-5.6 Sol in Mixedbread’s own search evaluations at up to 10x lower cost and 12x faster; available now via the Mixedbread API. (Mixedbread)
  • Alibaba Qwen: Released Qwen3.8-27B, a compact dense multimodal model distilled from Qwen3.8-Max with native vision-language support, 262k context, and adjustable reasoning depth, open-sourced under Apache 2.0 as a local-deployment alternative to the larger Qwen3.8-Max. (Hugging Face)
  • Google DeepMind: Released Gemini 3.7 Flash, its most intelligent workhorse model yet for coding and agents, shipping three weeks after Gemini 3.6 Flash with substantial coding gains (FrontierCode 43.6% vs 34.4%, DeepSWE 65.3% vs 49.0%) and an introductory price of $0.75/$3.75 per 1M tokens, half the prior Flash cost. (Google DeepMind)
  • DeepSeek: Open-sourced DeepSeek Harness v0.1 under the MIT license, a developer-preview agent framework built on a modular "everything-is-a-plugin" Cordis architecture that turns language models into autonomous coding/tool-use agents, positioned as an alternative to Codex and Claude Code. (DeepSeek)
  • xAI: Released Grok 4.6, a post-training upgrade to Grok 4.5 with a 500K-token context window, a new "xhigh" reasoning-effort level, and stronger long-horizon agentic/coding performance, now live via the xAI API, Cursor, and Grok Build at unchanged pricing. (xAI)
  • Alibaba Qwen: Released open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max) on Hugging Face, its largest open-weight model at 2.4 trillion total / 95B active parameters with a hybrid full/linear attention architecture and up to 1M-token context, moving the model from proprietary API-only access to publicly downloadable weights. (NVIDIA Developer Blog)
  • Suno: Launched Suno Studio 2.0, turning its AI music platform into a chat-driven DAW for Premier subscribers with MIDI import/recording/editing, stem separation, automation curves, a text-driven plugin/instrument creation chat feature, sidechain compression, convolution reverb, and a new wavetable synth. (Suno)
  • NVIDIA: Released NeMo Switchyard, an open-source model-routing library that directs each step of an agent workflow to the most capable/efficient available model, released alongside Nemotron 3.5 Lightning to cut cost on high-volume execution tasks like tool calls and result validation. (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.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive