In partnership with
In Today’s Issue:
🔐 OpenAI says Astra crosses its Critical cyber line
🧠 Astra's reasoning moves out of human view
⚡ Qwen's point release nearly triples its coding scores
🧬 Fable 5.1 designs proteins and remaps Venus
📊 Qwen's scoreboard against Opus 5 and Fable 5
🔬 An AI designed the experiment and the lab ran it
✨ And more AI goodness…
⚡ The Signal
Two of the leading labs spent yesterday putting two different locks on their best work.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1: one model, published twice, with Mythos reachable only by vetted cybersecurity and life-sciences professionals. The same day, OpenAI said its forthcoming Astra model is the first it has ever rated Critical for cybersecurity under its Preparedness Framework, and reporting since has centered on a training technique that makes part of Astra's reasoning unreadable to humans. One lab is rationing who may ask. The other is rationing how much of the answer you get to watch it produce. Both moved because their models crossed into work that counts outside the chat window: molecules that hold up in a laboratory, and software flaws no human had ever found.
All the best,
Kim Isenberg
(Art: Shane Burke; photo: Al Drago / Getty Images, via The Information)
🔐 OpenAI Says Astra Crosses Its Critical Cyber Line
OpenAI now rates its forthcoming Astra model at the Critical cybersecurity threshold of its Preparedness Framework, the first model it has ever placed there. The company says Astra can find previously unknown flaws and build working exploits across hardened systems without a person guiding each step: it scored 100% on ExploitBench, and on an internal set of recently disclosed browser vulnerabilities it discovered and chained two zero-days, which OpenAI says it is now disclosing to the maintainers. OpenAI notes both figures were measured with privileged Daybreak Blue access, above the default production configuration. It says it plans to make Astra available soon, with the strongest cyber capabilities going first to a small group of alpha testers.
👉 tl;dr: The first model its own maker calls dangerous enough at hacking that the best parts ship behind a waiting list.
(The Information)
🧠 Astra Does Its Thinking Where Nobody Can Read It
The industry's main safety tool for frontier models is reading what they write while they think, and Astra's new training makes part of that writing disappear. The technique, called recurrent depth, cycles text through the same layers over and over inside the network instead of spelling each step out in words. It is efficient, letting a smaller model perform like a much bigger one, and it removes precisely what investigators used to reconstruct July's Hugging Face hack: the written chain of thought in which agents admitted they were out of bounds and kept going anyway. OpenAI says it limited the technique so Astra stays legible. The worry is the labs that will not.
👉 tl;dr: No major commercial model has shipped with this before, and it moves part of the reasoning somewhere humans cannot follow.
(Chart: Superintelligence, from Alibaba Qwen's launch table)
⚡ Qwen's Point Release Nearly Tripled Its Coding Scores
Alibaba did not give this one a new version number. Qwen3.8-Max-0902 is a date stamp on the flagship it shipped a month ago, and on Qwen's own numbers it moves further than most companies manage in a full release: agentic terminal coding from 11.3 to 29.0, black-box software replication from 10.5 to 28.0, professional job tasks from 53.4 to 64.0. The parameter count and the context window are unchanged, at 2.4 trillion and one million tokens; Alibaba credits further post-training on coding and office work. It still lists at $2 and $6 per million tokens.
👉 tl;dr: The gains that used to require a whole new model now arrive as a dated point release.
🎬 Watch This
Anthropic's Model Hardware Standard is a proposal for letting AI agents safely operate physical equipment in scientific research and advanced manufacturing. The eleven-minute walkthrough explains why plugging a model into real laboratory instruments turns out to be largely a safety-engineering problem. It is the plumbing underneath every claim in today's issue about models starting to do science.
"I want to prevent a race into unmonitorability kicked off by confused reporting."
– Jakub Pachocki, Chief Scientist, OpenAI
Pachocki posted this after reports that Astra's training obscures part of its reasoning, adding that the depth of OpenAI's frontier models, Astra included, sits "within a factor of two of GPT-4."
Elon Musk says Grok 4.7 arrives in ten days, answering Shopify CEO Tobi Lütke's praise for Grok 4.6. Reported timelines have already moved once, from late August into September, which puts the middle of the month in play.
(Grok 4.6 against Grok 4.5, GPT-5.6 Sol and Fable 5. Compiled by explainx from published and self-reported scores)
Two Models, Two New Gates: Fable 5.1 Ships as Astra Stops Showing Its Work
The Takeaway
👉 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1: one model shipped twice, with Mythos reachable only by vetted cybersecurity and life-sciences organizations.
👉 Fable 5.1 costs about 25% less than Fable 5 on typical workloads and up to 45% less on agentic tasks, after a 75% cut to cache reads, to $0.25 per million tokens.
👉 Anthropic says Mythos 5.1 designed protein binders that bound ten times more tightly than the best entries in Adaptyv Bio's competitions, and that Fable 5.1 remapped a third of Venus at two to three kilometer resolution.
👉 The same day, OpenAI rated its forthcoming Astra model Critical for cybersecurity and said it will ship with additional chain-of-thought monitoring.
The industry stopped shipping one model for everyone this week. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, one model published twice under different safeguards: Fable for general use, Mythos reachable only through trusted-access programs for vetted cybersecurity and life-sciences work. The same day, OpenAI drew its own line, declaring the forthcoming Astra the first model to reach the Critical cybersecurity threshold under its Preparedness Framework. Two labs, one day, two admissions that their best work can no longer go out to everybody on the same terms.
What earns Fable 5.1 that treatment is the science. Anthropic gave Mythos 5.1 open-source protein design tools and sent its output to two external organizations for laboratory validation; on three targets, the binders it designed bound ten times more tightly than the best entries submitted to Adaptyv Bio's design competitions. Fable 5.1 trained a neural network on 30-year-old Magellan radar data and produced a new elevation map for a third of Venus, resolving features down to two to three kilometers instead of 10 to 20, and released it under a Creative Commons license ahead of NASA's VERITAS and ESA's EnVision missions. Anthropic calls this "an early glimpse of how AI models will contribute to scientific progress."
(One of Mythos 5.1's designed binders, orange, docked to the Nipah virus G protein. Anthropic)
The benchmark that moved most is Terminal-Bench-Science, which tests running real scientific work from a command line. Fable 5.1 more than doubles its predecessor there, 52.6% against Fable 5's 24.7%, and clears Opus 5's 29.0%. It got cheaper in the same release: about 25% less for typical workloads and up to 45% less on long agentic runs.
(Chart: Superintelligence, from Anthropic's published scores)
OpenAI's gate is shaped differently. Astra earned that Critical rating on the strength of its hacking, including a perfect score on ExploitBench and two zero-days it found and chained itself. The complication, reported by The Information, is that part of Astra's gain comes from recurrent depth, a looped design that moves some of the thinking inside the network, where there is no written chain of thought left for anyone to read. OpenAI says it limited the technique so Astra's reasoning stays legible, and chief scientist Jakub Pachocki warned publicly against "a race into unmonitorability."
None of this stayed on a screen. A laboratory built the binders and measured them, and the Venus map is finished and free for anyone to download today. That is the real change here. Until now these models helped scientists do their work. This week one of them did a piece of the work itself and handed over the result.
Why it matters: Two of the leading labs decided on the same day that their best model cannot simply be handed over on the old terms. Access, rather than raw capability, is becoming the thing that gets rationed.
Sources:
🔗 https://www.anthropic.com/claude-fable-and-mythos-5-1
🔗 https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235
🔗 https://openai.com/index/path-to-astra/
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:
7 high-impact AI strategies to accelerate your marketing performance
Practical use cases for content creation, lead gen, and personalization
Expert insights into how top marketers are using AI today
A framework to evaluate and implement AI tools efficiently
Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.
The chart: Qwen's launch table for Qwen3.8-Max-0902, scoring it against Claude Opus 5, Fable 5 and GPT-5.6 Sol across coding, agent and multimodal tests. Against the American frontier it splits the board: it takes repository-level code understanding (66.3 to Opus 5's 63.2), machine-learning research engineering and the visual reasoning rows, and it loses most of the pure coding tests, including 29.0 to Opus 5's 42.7 on TerminalBench 3.0.
The lesson: Across roughly half of this table a model listed at $2 and $6 per million tokens finishes within a few points of models listed at $10 and $50. The capability gap is now far smaller than the price gap.
The caveat: These are the vendor's own numbers, and three of the benchmarks (QwenSWEBench V2, CoWorkBench and WorkArena) are Qwen's in-house tests. The comparison runs against Fable 5 and Opus 5, not the Fable 5.1 Anthropic shipped the day before, and Qwen's own footnote warns that its Fable 5 results "may involve fallbacks."
🔬 The Week an AI Designed the Experiment and the Lab Ran It
⚡ Bottom line: Google DeepMind's Co-Scientist designed real experiments in three fields, and human labs carried them out successfully.
💡 Why it matters: Earlier tests had AI agents writing papers. This one has them producing physical materials and verified predictions.
🔎 What it means: The bottleneck moves from generating hypotheses to running enough real experiments to test them.
The experiment in one sentence: give an AI system a real laboratory problem, let it design the recipe, then have human scientists actually run it and see whether the material comes out. A team of 35 researchers led by Google DeepMind's Samuel Schmidgall, with collaborators at Duke and Columbia, did that across three fields and posted the results on August 27.
(How Co-Scientist works, and the three fields it was tested in. Schmidgall et al., arXiv)
In materials science the target was an MXene, a two-dimensional material normally made using hydrofluoric acid. Co-Scientist was told what the lab could and could not do, and asked for a safer route. It proposed a chemical vapor deposition recipe built on hexachloroethane and titanium; the lab ran it, and the material that came out carried the key structural signatures of the Ti3C2Tx MXene lattice. It also produced growth recipes for three monolayer semiconductors that worked on the first attempt.
(The vapor deposition route Co-Scientist designed, and the layered material it produced. Schmidgall et al., arXiv)
In biology it was handed sparse images of engineered E. coli and asked to predict how the colonies would spread across different chemical concentrations. Its predictions matched laboratory measurements that had never been published. In computer science it designed its own inference-time architecture for medical response generation, which beat six frontier models on HealthBench.
A month ago the question was whether an AI agent could write a paper worth publishing. This week a machine designed a synthesis route, a human lab ran it, and a new two-dimensional material came out of the tube. It pulled off the equivalent in two more fields: bacterial colonies predicted before the lab had published the measurements, and a medical-response architecture that beat six frontier models on HealthBench. The question now is whether laboratories can run experiments fast enough to keep up with the ideas coming out of the machine.
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
