🧪 Anthropic’s Week of Intentions: A Wet Lab, Self-Measured R&D, and Claude Leading 26% of Its Own Research

Three Anthropic stories landed inside 24 hours and they read as one argument.

First, the lab. Reuters reported on Sep 18 that Anthropic has quietly built a wet lab — a facility for physical experiments — in the San Francisco Bay Area, moving its biology work well past in silico evaluation. Anthropic’s head of life sciences, Eric Kauderer-Abrams, confirmed the lab and said the ambition is for Claude to direct robotic units with limited human intervention, while insisting that human oversight remains essential. The company has spent 2026 assembling this: the Claude Science software, Novartis CEO Vas Narasimhan joining the board, and the roughly $400M stock acquisition of Coefficient Bio. The boundary Anthropic draws is that it is not running clinical trials, so it does not compete with pharma. Isomorphic Labs, Alphabet’s drug-discovery arm, has worked on the clinic for years and has already slipped its own target to the end of 2026.

Second, the measurement. On Sep 17 Anthropic published three metrics for tracking how fast AI development is moving: how much AI R&D work Claude leads, how agent actions are overseen, and how compute is allocated. The numbers it disclosed are unusually specific. As of August 2026, Claude leads 26% of Anthropic’s AI R&D work, up from under 1% in February 2026, and is at or above the collaborates level on more than 90% of it. Roughly 30,000 agents were doing research and engineering work at any one time on its most-used internal platform. Online monitors check 100% of agent actions before execution, usually within seconds — and of more than a billion agent decisions analysed in August, 0.002% (about 1 in 47,000) were blocked. About 6% of research compute went to safety in a July 13-20 snapshot. Anthropic states plainly that Claude is not fully autonomous in any measured subset of its R&D work.

Third, the turn. Claude now leads roughly a quarter of the work that builds the next generation of Claude models. That flywheel is exactly what Dario Amodei’s coordinated-slowdown plan is aimed at, and this post is its missing operating manual. The stated reason for publishing is a good one: minimise the gap between what frontier labs know and what the public knows. The obvious limitation is also worth stating: these are self-reported numbers, produced by the lab being measured, on a frozen task basket that by Anthropic’s own admission captures the automation of existing work rather than the arrival of new kinds of work, and they cover one internal platform only.

⚖️ Unsealed Filings: Microsoft’s Own Team Called AI Scraping the Largest Theft of Labor in Human History

Newly unsealed documents in The New York Times’ three-year-old copyright suit against OpenAI and Microsoft contain the kind of quotes defendants normally prefer to keep sealed. Much of the material comes from the Times’ brief rather than the underlying exhibits, which remain sealed — but the language is blunt.

A January 2023 internal memo by Brent Hecht, Microsoft’s director of Applied Science, called the companies’ data acquisition an astonishing theft of unprecedented proportions and the largest theft of labor in human history. A January 2024 Microsoft presentation described the collapse in click-through rates to nytimes.com caused by Copilot’s answer engine — as much as 93% versus traditional Bing search — as a doom loop that would hurt the performance of our models and the entire web at the same time. The same document concedes that it is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its content supply chain.

Other admissions target different pillars of the fair-use test. OpenAI’s head of ChatGPT, Nick Turley, wrote internally that publishers face an existential threat from a product that is largely substitutive and will get more and more substitutive as they get better. Satya Nadella testified earlier this year that anything that is paywalled should be licensed by anyone who wants to use it for grounding or training, and said he would have invoked Microsoft’s right to require OpenAI to retrain its models had he known paywalled material was used. When OpenAI researcher Nick Ryder described a hack to get around nytimes paywall, Greg Brockman replied ah nice.

The scale is the part that is new. OpenAI’s mid-training datasets alone contain more than 91,692 copies of works published by the Times, the Daily News and the Center for Investigative Reporting; a Common Crawl-derived dataset held more than 2 million documents from nytimes.com alone; and data routed through Microsoft initiatives called Project Taxi and Project Mango contains at least 160,903 unique works from the news publishers. Microsoft disavowed the internal language: Microsoft’s position is set out in its court filings, which explain why these transformative uses are consistent with copyright law and why Copilot is not a substitute for publishers’ journalism.

Fair use turns on whether the use harms the market for the original work. Internal documents that predict that harm, measure it, and describe it as an existential threat are the hardest kind of evidence to argue around.

🏗️ Compute Splits in Two: Crusoe’s $3.9B and the Hunt for Smaller Data Center Deals

Crusoe raised $3.9B at a $30.9B valuation to build both very large data centers and small modular AI factories — capacity sited near power rather than merged into one campus. The same day, CNBC reported that Anthropic and OpenAI are hunting smaller data center deals, a visible shift from the gigawatt-scale announcements of the spring. And a Senate bill intended to curb data-center electricity costs hit a snag, one day after the House passed the Ratepayer Protection Act 417-3.

The buildout is not slowing; it is fragmenting. Smaller sites come online faster, dodge a single grid-interconnection queue, and generate less ratepayer backlash than one 400 MW campus in one state. If the constraint of 2025 was capital, the constraint of 2026 is power and politics — and the shape of the deals is starting to show it.

🛫 The FAA Bets $875M on AI to Untangle Air Traffic

The Federal Aviation Administration is preparing to launch SMART — Strategic Management of Airspace, Routes, and Trajectories — an AI software program meant to help air traffic controllers manage their workload and route flights more safely. It costs $875M over 12 years and comes from a company called Air Space Intelligence. According to the FAA’s own one-pager, SMART uses AI to assess airline schedules, weather, airport capacity, airspace conditions and operational constraints to predict traffic flows and identify potential conflicts before they occur. Rollout starts in the Washington, D.C. metropolitan area.

This is the least glamorous and most consequential AI deployment of the cycle. It is a safety-critical system in a sector with a chronic controller shortage, and it is the first place in this series where an AI system’s errors will be counted in lives rather than click-through rates. That is the reason the twelve-year price tag may prove the least interesting number in the contract.

🇺🇸 Election-Season AI Politics: States Defy Trump as AI Titans Get a Seat at the Trump-Xi Table

  • States are not waiting. Politico reports that Republican and Democratic states alike are proceeding with their own AI rules despite the White House’s deregulatory push, after the House passed the Ratepayer Protection Act 417-3.
  • AI executives are on the state-dinner guest list. Sam Altman and Jensen Huang are reportedly planning to attend the White House dinner for Xi Jinping’s visit, with Qualcomm’s Cristiano Amon also expected around the Sep 24 summit; AI safety, development standards and chips are the agenda.
  • Liability is the industry’s new worry. The New York Times reports that as agents start acting rather than answering, the question shifting to the front is who is responsible for a bad action.
  • OpenAI went after law firms, launching a legal-focused AI platform in a race for legal-sector users.
  • Forecasting got a funding round. London-based Mantic raised $25M in seed funding led by Radical Ventures, with Microsoft’s M12, Thinking Machines Lab and Balderton Capital participating, after outperforming every human contestant in the summer 2026 Metaculus Cup by assigning calibrated probabilities to political, economic and cultural events.
  • Also in the window: Google is helping draft state chatbot-safety laws while lobbying for carve-outs; MIT’s AI committee report on cognitive surrender drew wide coverage as universities struggle to write consistent rules; a grandmother wrongly accused of bank theft after an AI facial-recognition match sued for $10M; the WSJ detailed a Chinese hacking firm using AI to supercharge cyber-espionage; and Al Jazeera reported on ISIL using commercial AI to design explosives.

🧩 Typed Decisions Go Mainstream: the Jev Ecosystem Passes 400 Repositories in 48 Hours

GitHub’s search API returns 403 repositories created since Sep 17 that mention Jev — the TypeSafe model whose pitch is that it returns typed decisions with calibrated probabilities instead of generating tokens. Two days after a public launch, the derivatives cover browser automation, shell history, semantic HTTP routing, a context-pruning proxy for Claude Code and Codex, a vLLM adapter, a Rust linter, a Chrome ad-blocker and a dozen curated awesome- indexes. Yesterday’s issue covered the fastest of them; today the interesting signal is the shape of the crowd.

The same instinct showed up outside the Jev orbit. NandhaKishorM/laya appeared the same day — a non-autoregressive System 1 decision engine with mathematically calibrated probabilities, shipping with a PyPI package, a Hugging Face model and a Colab notebook. On Hacker News, Bend — a language that uses proofs rather than prompts to block AI mistakes, with a CPU and GPU backend — hit 510 points.

The architectural move underneath all of it is the same, and it is vendor-neutral: pull a cheap, verifiable classification step out of the token-generation loop, and let the generative model do only the work that genuinely requires generation. It is the same reasoning behind the sub-cent-per-step computer-use repos from yesterday’s issue, and it is the first design pattern the agent era has produced on its own terms rather than borrowed from LLM serving.

⏰ OpenRouter Limited-Time Free Models

Model Context Price Expires
Dots Studio: Dots3-Note Preview (free) 512K Free 2026-09-30 (11 days left)

Only one limited-time promotion is still live, and it is the same one as yesterday — nothing new was added to the promo shelf this cycle. The permanent :free shelf on OpenRouter remains the larger free tier; it is a standing free offering rather than a promotion, and it is where labs quietly park models while they are being evaluated.

🆕 New OpenRouter Models

Model Created (UTC) Context Price
unbiased/pareto 2026-09-17 23:02 262K $2.50/M tokens in, $7.50/M tokens out (cache read $0.25/M tokens)

Pareto is described by OpenRouter as a multimodal composite model built for research, coding, and agentic workflows, while delivering frontier-level performance across a broad range of general-purpose tasks — text and image input, a 262K context window and up to 131K output tokens, with tool calling and structured outputs supported. A composite model that carries a 3x input-to-output price ratio and a 10x-cheaper cache read is worth noting for a specific reason: composite and routed models are priced like products, not like single checkpoints, and their real cost depends on which sub-model ends up serving the request.

It is the only listing created since our last issue. The check that established that is the created-timestamp window — every other recent OpenRouter entry, including the DeepSeek -latest aliases and the OpenAI Astra/Sol/Terra/Luna family, predates the previous publish.

🤗 Trending on Hugging Face

The collector’s raw new-uploads list was again dominated by zero-download test repos this cycle — the fourteenth consecutive occurrence — so this section uses Hugging Face’s official trending feed: recently popular rather than newest.

Model Downloads Likes Type
Qwen/Qwen3.8-27B 7.36M 15,599 image-text-to-text
Lightricks/LTX-2.5 1.59M 4,277 image-to-video
ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF 1.08M 1,290 quantized
deepseek-ai/DeepSeek-V4.1-Flash 429,865 3,100 image-text-to-text
prism-ml/Ternary-Bonsai-2-27B-gguf 405,609 642 ternary / quantized
Edge0/Edge0-35B-A3B-preview 52,519 3,366 text generation (MoE preview)
TokenRhythm/NeoHorse-1-4B 22,666 2,359 text generation
openbmb/UltraData-SFT-Agent-2609 16,384 190 agent SFT dataset
m-a-p/YuE2-3B 13,668 767 text-to-audio
markov-ai/cad-1000-hours 153,590 471 CAD dataset

The composition of this feed is the story. Qwen3.8-27B is now a family rather than a model — the base checkpoint plus quantization and fine-tune derivatives past 2M downloads, with a new community derivative (ukisai/Swift-Qwen3.8-27b) appearing today. Second, compression is the live research front: a ternary 27B GGUF and a 4-bit-class GSQ variant both sit in the top five, which lines up with the behavioral-ablation tooling that showed up on GitHub today. Third, Edge0’s preview still carries 3,366 likes on 52K downloads — a like-to-download ratio of 1:16, the signature of a model people are evaluating rather than deploying.

⭐ GitHub Trending: AI Edition

All five entries in the collector’s list were created between Sep 11 and Sep 16 — before our last issue — so the entries below were substituted from GitHub’s search API and each creation timestamp was verified individually.

Repo Stars What it is
Continuum-AI-Corp/OrcaBonsai-27B-Uncensored ⭐ 136 Runtime behavioral ablation for compressed LLMs — no weight modification, no re-quantization. First target: Ternary Bonsai 2 27B
NandhaKishorM/laya ⭐ 56 Fast, non-autoregressive System 1 decision engine with mathematically calibrated probabilities; ships with a PyPI package, HF model and Colab notebook
indada/repopilot ⭐ 55 Verification-driven AI software iteration on the OpenAI Codex SDK — self-hosted agents turn goals, issues and PR feedback into tested, reviewable changes, with merging and deployment kept under human control
VDERR/echocat-skill-panel-3.0 ⭐ 53 In-app skill manager plus an audit trail for agent skill invocations
gylive/ccodex-sleep-state ⭐ 52 Collects and re-injects turn-state locally to work around Codex quality throttling — a sign that subscribers are now debugging their coding agent’s infrastructure themselves

Worth naming alongside them: cobanov/awesome-jev (⭐ 42) is one of a dozen awesome-Jev indexes that appeared today, and it is the cleanest illustration of what happens when a new primitive ships into an ecosystem that has spent two years building prompt-and-tool wrappers.

💡 Key Trends

  1. Measurement is becoming the product, and the scoring problem comes with it. Anthropic published its own R&D automation index and OpenAI now runs a standing disclosure framework. Both are real steps: the industry is producing, for the first time, numbers a policy argument can be had about. Both are also self-reported, on self-chosen baskets, by the labs whose pace is in question. The next useful thing is not another index — it is an audit.
  2. The buildout is fragmenting rather than slowing. A $3.9B round into small modular AI factories, two frontier labs hunting smaller data-center deals, and a Senate fight over who pays for grid upgrades. Capital is abundant; power, permitting and politics are the binding constraints, and that is reshaping what gets built rather than whether it gets built.
  3. Oversight statistics can be read two ways, and both readings matter. Anthropic’s monitors checked more than a billion agent decisions in August and blocked about 1 in 47,000. Either the agents almost always behave, or the checks are shallow enough that they rarely trip. The same week, the agent ecosystem on GitHub grew by 400 repositories in 48 hours — adoption and control are still visibly different curves.
  4. AI is moving into physical and safety-critical work, which changes the evaluation question. An $875M air-traffic platform, a wet lab where Claude is meant to direct robots, a robot model shipping zero-shot home generalization. In each case the metric that matters stops being benchmark accuracy and becomes what happens on the day the system is wrong.

Today’s Snapshot

  • Coverage window: 2026-09-17 13:07 UTC → 2026-09-18 13:03 UTC
  • New OpenRouter models: 1 (unbiased/pareto, created 2026-09-17 23:02 UTC)
  • Limited-time free models: 1 (Dots3-Note Preview, expires 2026-09-30)
  • Hugging Face trending models: 10 collected, 10 listed
  • GitHub repos: 5 (API-verified substitutes for a 100% out-of-window collector list)
  • arXiv papers: 0 collected this cycle
  • Collector total: 17 items
  • Dominant story: Anthropic’s three-part push — the wet lab, the R&D metrics, and Claude leading 26% of its own research — closely followed by the unsealed Microsoft and OpenAI filings in the Times copyright case
  • Notable single-day numbers: 26% of AI R&D work led by Claude, up from under 1% in February; 0.002% of over a billion agent actions blocked; 93% click-through decline to nytimes.com; $3.9B raised at a $30.9B valuation; $875M over 12 years for SMART; 403 Jev repositories created in two days