Astra hits a Critical cyber bar + Anthropic's Fable 5.1

OpenAI now classifies its unreleased Astra model at the Critical cybersecurity level of its Preparedness Framework — the first model it has ever placed there. In expert-led testing, Astra found previously unknown vulnerabilities in a hardened browser and chained them into a working sandbox escape that ran commands on the host, and it scored a perfect 100 percent on a public exploit benchmark. OpenAI says it delayed parts of Astra's development for several weeks, restarted the paused frontier training run on August 28, and will limit advanced cybersecurity access to a small group of alpha testers at launch.

Anthropic released Claude Fable 5.1 and its less-restricted sibling Mythos 5.1, reporting 55.8 percent on the Terminal-Bench 4.0 coding benchmark against Fable 5's 42.0 percent. The headline per-token price did not move — it is still $10 input and $50 output per million — so the savings Anthropic estimates, roughly 25 percent on typical work and 45 percent on agentic work, come from cutting cached reads to 25 cents per million and from reducing false-positive refusals by 60 percent on cybersecurity and 85 percent on biology. Mythos 5.1 goes only to registered partners doing cybersecurity or life-sciences research, the same vetted-access approach that OpenAI applied to Astra on the same day.

Apple filed new material in the Northern District of California accusing Chang Liu, a former Apple engineer now working at OpenAI, of using a confidential Apple circuit schematic and an internal engineering tool in his work there. The filing further alleges that Liu enlisted a colleague to destroy evidence in June after learning of the investigation. Apple is seeking a preliminary injunction blocking OpenAI hardware development that draws on its technology; OpenAI calls the access routine residual access that Apple failed to revoke, and notes that more than 400 former Apple employees now work there.

OpenAI opened an Epic integration that pulls authorized patient context — visit notes, laboratory results, medications and specialist documentation — into ChatGPT for Healthcare, so a clinician can ask what changed since a patient's last visit instead of reading through the chart. A companion plugin adds structured access to nine official public sources including PubMed, ClinicalTrials.gov, DailyMed and Medicare coverage policy. The University of California, San Francisco health system is a pilot partner. OpenAI says physicians across 60 countries have reviewed more than 700,000 model responses to tune its healthcare behavior.

Google Research and NASA's Jet Propulsion Laboratory published a deep-learning system that finds, measures and traces methane plumes in imaging-spectrometer data from an instrument aboard the International Space Station. It reaches 84 percent recall against expert-annotated plumes, surfaced about 50 percent more plausible plumes across roughly 1,100 scenes, and mapped emissions at 24 of the world's 25 largest-emitting landfills at 60-meter resolution. The plume database is published on Google Earth Engine, and the trained model and the synthetic training data are on Kaggle.

World Labs, the spatial-intelligence company co-founded by Fei-Fei Li, introduced Atlas, a single multimodal model that generates camera-controlled video running up to one minute at 1440p, reconstructs three-dimensional scenes from images, and generates images from text. The company reports 81 percent human preference over Gemini Omni Flash on camera-controlled generation and lower reconstruction error than specialist models on several standard datasets. Atlas is entering early access with selected partners through a waitlist, with no open weights, no public interface for developers and no published pricing.

Google DeepMind shipped agentic video understanding for Gemini, which lets the model search, scan and re-inspect specific segments of a video across frames, audio and transcript rather than sampling at a fixed frame rate. Google reports up to 88 percent fewer tokens, up to 66 percent lower cost and up to 7 percent better accuracy on video tasks. It is live now in the Gemini application programming interface through Google AI Studio and the Gemini Enterprise Agent Platform, with the Gemini app and a YouTube question-answering feature to follow.

At the Farm Progress Show, John Deere introduced JD, a generative AI assistant inside its Operations Center that answers questions about a farm's own planting, spraying and harvest data — how fuel use compares across three seasons, which operator covered the most acres, when the data says harvest should start. It opened as a limited early-access program for selected US customers, with web, mobile and in-cab availability planned later in 2026, and it is free for app users. Deere published a ten-point farmer data commitment alongside it, which promises that growers own their data and that the company never sells it.