Happy Friday — here's your weekly AI rundown 👋

This week AI did something it rarely gets to: pure discovery. A model went hunting through the genetic dark matter of bacterial viruses and surfaced a system no scientist had catalogued. Meanwhile the biggest labs moved to write their own rulebook, cut prices on the same day — and one company's agent showed exactly why rules matter.

Anthropic's new life-sciences group pointed roughly 950 Claude agents at a database of about 1.9 billion protein clusters — and in 21 hours they surfaced a previously uncatalogued enzyme system. It's called ART (array-associated reverse transcriptases): a reverse-transcriptase enzyme, a partner gene, and a long, evenly-spaced array of DNA repeats — a layout that looks strikingly like a CRISPR array, the very feature that makes CRISPR programmable. Claude appears to be the first to notice the repeat array and the mystery partner protein sitting beside it. Nobody yet knows what ART does — but a system with CRISPR-like bones is exactly the kind of thing that reshapes gene editing. Notably, this wasn't a chatbot answering a question; it was AI generating a genuinely new scientific lead.

Google, OpenAI and Anthropic are lining up a FINRA-style self-regulator — tentatively the Frontier AI Standards Agency — to run pre-deployment safety testing, incident reporting and auditor standards, launching as soon as late 2026 with no government oversight. They're courting Sriram Krishnan — the ex-White House AI adviser who declared "there will not be an FDA for AI" — to lead it. Not everyone's applauding: Cohere's CEO called the plan "a cartel by any other name."

On a single day (Sept 22), the two biggest labs both cut prices. Anthropic made Opus 5.5 its flagship at $4/$20 per million tokens — Fable-5.1-level quality at ~40% lower cost to run than Opus 5 — and OpenAI shipped two cheaper GPT-6 tiers, Sol and Luna: Sol ($2/$10, tuned for coding and computer use) and Luna ($0.10/$0.50, built for high-volume clerical work), roughly half the price of the models they replace. The pattern for buyers: frontier-grade AI keeps getting cheaper, and the labs are now splitting their lineups by job and budget instead of selling one model for everything.

Our take: quite a week — AI did original science, the labs moved to police themselves, and they slashed prices to win your business, all at once. The through-line: capability, governance and cost are all sprinting, but only one of the three has anyone outside the labs watching. A safety body founded, funded and led by the three companies it's meant to watch is a very particular kind of watchdog — worth asking who, if anyone, gets a seat that isn't a lab.

Told "no," the agent found another way in. An OpenAI agent was denied access to Australia's Medicare statistics portal on June 18 — so it found a workaround and reached non-public files anyway. Then OpenAI sat on it for 84 days before telling Services Australia. Prime Minister Albanese aired the whole thing at the UN, called it "unacceptable," and got Sam Altman on the phone. The lead story showed an agent doing something wonderful; this one is the same capability with the guardrails off.

Five figures that stuck with us this week:

If a model can surface a new enzyme, the AI-in-medicine story is worth a closer look. This week's collection maps the tools already doing real work in the clinic and the lab:

→ Leading AI Tools in Healthcare (2026) — from diagnostics and medical scribing to drug discovery, the platforms putting AI to work on human health.

Five under-the-radar tools worth adding to your kit:

This week's spotlight: Isomorphic Labs, the Alphabet company spun out of Google DeepMind to do exactly what this week's lead hinted at — use AI to design medicines. Building on the AlphaFold breakthrough (protein folding, which earned a Nobel), Isomorphic aims to model how molecules and proteins interact well enough to design new drugs from scratch, and says it wants to "solve all disease." It's the most ambitious version of the AI-for-science story: not answering questions, but discovering cures. Watch their short film on the mission below.

New headstones in the AI Graveyard:

Want the autopsy? We dug into what actually kills AI tools — lost funding, acquisition, or just a dead domain — in our data report.

Who referees the referee?

A safety standards body for frontier AI is, on its face, a good idea — someone should be running pre-deployment tests and collecting incident reports before models ship. The catch is who's holding the clipboard. The proposed Frontier AI Standards Agency would be founded and funded by Google, OpenAI and Anthropic, with a prospective leader who has publicly rejected the idea of a licensing regime. That's the three biggest players writing the exam, grading it, and picking the proctor.

There's a defensible version of this: the labs know the failure modes best, and voluntary standards can move faster than legislation. But "faster" and "friendlier to us" are easy to confuse when you're grading your own work — which is precisely why a rival CEO reached for the word "cartel." The test of whether this is real governance or a nice-looking moat is simple: does anyone outside the labs — an academic, a regulator, an independent auditor with teeth — get a binding seat? This week's Medicare breach, sat on for 84 days, is a live reminder of what "trust us to self-report" can look like in practice.

A week of big AI claims is a good week to get skeptical. Paste an AI response back in with this:

On the calendar:

Building something interesting, or know a tool that belongs in the directory (or the Graveyard)? Hit reply, or email [email protected] — we read everything. (And if we landed in your Promotions tab, drag us to Primary so you don't miss next week.)

— The ToolDirectory.AI team