Anthropic shipped one model under two names and two safety settings
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026 -- the same underlying model shipped twice, with the only difference being how tightly its cybersecurity and biology safeguards are wound.
Claude designed protein binders that worked about half the time
Anthropic gave Claude Mythos 5.1 open-source protein design tools and sent its output to outside labs, where nearly 50 percent of its designs across 12 targets bound successfully -- against the 10 to 15 percent hit rate typical of the field.
OpenAI formally designates Astra as its first Critical cyber-capability model
OpenAI announced on September 1, 2026 that its Astra model meets the Critical cybersecurity threshold under its Preparedness Framework -- the first model the company has ever placed at that level -- after experts used it to find unknown browser and operating-system vulnerabilities and chain two zero-days into a working exploit.
CrowdStrike shipped an attacker model and a defender model that train against each other
CrowdStrike launched SafeMind on September 1, 2026 -- a pair of security models built on NVIDIA's Nemotron, one offensive and one defensive, run in a closed loop where each is continuously pitted against the other to improve.
Anthropic closed the hole distillers used to read Claude's thinking
With Claude Fable 5.1, Anthropic blocked new API accounts from editing earlier turns of a conversation while keeping Claude's prior reasoning in the transcript -- shutting off a publicly documented technique for extracting a model's internal thinking at scale.
A 104 GB model now runs on a 48 GB Mac by streaming experts off the SSD
slotstream, a single Swift binary released as a Show HN on September 1, 2026, runs the 104 GB Qwen3.8-Flash-Next mixture-of-experts model on Macs with a fraction of that memory by keeping a small trunk resident and reading expert weights off the SSD on demand -- about 12 tokens per second on a 48 GB machine.
World Labs' Atlas generates a minute of 1440p video you can actually steer
World Labs introduced Atlas on September 1, 2026, a world model pretrained from scratch on text, images, video and 3D that grounds every input image at a position in space, letting it generate up to a minute of 1440p video with exact camera control instead of text-prompt guesswork.
Dan Luu scored every dated Ed Zitron AI prediction. All of them came back wrong.
Dan Luu published an audit on September 1, 2026 of roughly 28 dated predictions by the AI skeptic Ed Zitron going back to February 2024, and marked every single one wrong -- including repeated calls that generative AI had permanently peaked, that OpenAI's growth was collapsing, and that Cursor would die.
DoltLite hit beta on about 2,000 agent-written pull requests
DoltHub announced on August 31, 2026 that DoltLite -- a SQLite fork with Git-style version control -- reached beta after five months and roughly 2,000 pull requests written by a team of agents, and now passes 100 percent of the 5.8-million-query sqllogictest suite.
A 65,000-line Go-to-Rust rewrite cost $400 by translating through a data model first
Developer Iurii Krasnoshchok published an account on September 1, 2026 of rewriting a 65,000-line Go codebase into Rust for about $400 using Claude Fable, by having the model extract the program's structure into graphs and state machines first and regenerate code from that representation rather than translating file by file.
The Codex desktop app ships a full copy of LibreOffice
Simon Willison found on September 1, 2026 that OpenAI's Codex desktop app caches 1.7 GB of bundled runtimes -- a complete Python installation, a complete Node.js installation, and native binaries for git, Poppler and the entire LibreOffice office suite -- which the agent's skills then invoke to handle documents.