Claude gets cheaper and private; ChatGPT reads Epic charts
Edition of 2026-09-02
# AI For All ### Wednesday, September 2, 2026
As of today, a founder who pays for Claude by usage can cut the bill by about a quarter on typical work, and an enterprise can finally run it with complete privacy.
Cheaper, and finally private
Anthropic released Claude Fable 5.1 on Tuesday, and it is live everywhere: the API, Amazon, Google, and Microsoft clouds. For typical workloads the bill drops *25%. For long agentic tasks, the kind where the model works on its own for minutes without being told each step, the savings reach 45%*.
The other change is a privacy promise Anthropic could not make before. Its new Enterprise Frontier Safeguards stores everything on infrastructure the customer controls, not Anthropic's, with the same monitoring for misuse. Until that system arrives this fall, eligible customers get zero data retention from day one. The safeguards also stopped crying wolf: *60%* fewer false positives in cybersecurity.
Someone paying a monthly bill for Claude gets the same work for less, and a regulated company that could not touch the model over data rules now can.
The chart is now a question
ChatGPT Health now connects to Epic, the medical records system that holds data on over *325 million* patients. A clinician can pull in appointment notes, lab results, medications, and specialist reports, and ask questions about them. In some hospitals, ChatGPT appears directly inside the chart, for pre-visit reviews without leaving the patient's record.
The detail that matters is the lock: the integration is read-only, and the AI cannot write anything back. OpenAI says the setup was tested on responses reviewed by physicians and found *99.1% safe, a number worth reading as a strong signal, not a guarantee, given the stakes. People already ask 300 million* health questions of ChatGPT every week, so this is the same tool moving from the living room into the exam room.
For a clinician at an Epic hospital, the records just became one question away. For everyone else, it is the clearest sign yet that medical AI is being wired into the systems doctors actually use.
The model that breaks in is coming, on a leash
OpenAI says its next model, Astra, is the first to cross its Critical cybersecurity threshold. It can find unknown flaws in computer systems and exploit them without step-by-step guidance from a human. On the public hacking benchmark ExploitBench, OpenAI reports a configuration of Astra scored *100%*, and the same tests saw it find unknown flaws and weave them into an exploit chain.
OpenAI paused parts of Astra's development in August after preliminary evaluations, and now says it will arrive soon, with the strongest capabilities held back from the open market. The first access goes to Daybreak, a cybersecurity coalition of vetted organizations. No third party has independently confirmed the claims, and OpenAI has not said whether it is working with the US government on the evaluation.
For a security officer, the planning assumption just changed: assume a model that finds and exploits flaws on its own is months away, and that the people who get it first are the defenders, not the attackers.
One photo, any angle, one minute of video
World Labs, the company founded by Fei-Fei Li, opened early access to Atlas, a model that turns one or a few photos into video. You choose the camera position, and the scene keeps going: around a corner, behind an object, into parts of the room that were never in the picture. Up to *1 minute* of video, sharper than high definition.
The same model rebuilds scenes in three dimensions and simulates them over time, which is what robotics teams need for training. Atlas will power the next versions of Marble, World Labs' creative tool. It is early access, so nothing is usable at scale yet, and the demo reel is the vendor's own.
For a filmmaker or a visual-effects artist, a shot that used to need a rig, a reshoot, or a green screen is now something to request, and wait for.
Two arms you can order today, shipping this fall
Nori Robotics, out of this year's Y Combinator batch, opened orders for the Nori A3, a two-armed table robot, no deposit. It trains at home from a laptop, runs *6-8 hours* on a charge, and its skills are shared through a marketplace. Delivery is promised for this fall.
The arms lift about a kilogram and a half each, which puts it in the range of kitchen and workshop tasks rather than heavy labor. The orders are open today, with the full price and no deposit.
For a hobbyist or a robotics teacher, a two-armed robot you can train at home stopped being a lab-only purchase.
Weak signal
Google launched Google Pics, an image creation tool inside Docs and Slides, powered by its Nano Banana model, rolling out over the coming weeks to most Workspace business customers. No verifiable numbers on the primary page yet, so it stays a watch item: the moment Google puts usage or pricing figures behind it, it becomes a story.
The month opens with the workhorse getting cheaper, the exam room getting wired, and the most dangerous model yet announced on a leash. September will tell which of those holds.