Welcome back! Claude can now use your Mac while you work on something else. On the same day Google shipped a $0.75 model that beats Opus 5 on 9 of 16 tests, the New York Stock Exchange told Congress that Claude found massive vulnerabilities and Meta caught up to the frontier.
In today's Generative AI Newsletter:
Anthropic: What can Claude do on your Mac while you're not looking?
Google: Where does the cheap Gemini beat Opus 5 and where does it get crushed?
NYSE: What did Claude find inside the stock exchange?
Meta: Is the two-lab race over?
Claude got the biggest upgrade ever
Claude can now use your Mac in the background. Since March you had to watch it drive your screen. Now it clicks, types and opens apps on its own while you work in another window.
It doesn't touch your mouse or keyboard and waits if you're mid-typing.
It tries a connector first, then its built-in browser, then your screen only if there's no other way.
Trading and crypto apps are blocked by default and you can blocklist anything else.
It takes screenshots the whole time, so it sees everything on your screen. How comfortable you are with it depends on you. Beta, Pro and Max only, macOS 15 or later.
Turn it on in Settings → General → Computer use. OpenAI's Codex did this on the Mac first earlier this year. Anthropic finally matched it.
Google's $0.75 model beats Opus 5
Google just shipped Gemini 3.8 Flash , its third Flash in six weeks. Same price as 3.7 Flash, $0.75 per million input tokens and $3.75 output.
Google says it beats Claude Opus 5 and GPT-5.6 Sol on 9 of 16 benchmarks.
89.4% on Terminal-Bench 2.1, Opus 5 got 89.1%.
54.9% on HLE-Verified, best of the three.
19.1% on the harder Terminal-Bench 4.0, Opus 5 got 51.8%.
59% on OSWorld computer use, Opus 5 got 75.4%.
Google says it works harder on tough tasks, more reasoning steps and more tool calls. More tokens too, so watch your bill.
There's a Flash Cyber version for more than 650 vetted defenders. Chrome's security team says it wrote 2.6x more correct patches than the much bigger models.
It's live in the Gemini app for Pro and Ultra. For the price I'd throw every short task at it and keep Opus or Fable for the long agent runs.
Claude saved the stock exchange
NYSE president Lynn Martin told Congress yesterday that Claude Mythos found security holes in the exchange's systems and they fixed them fast.
A Michigan Republican asked how the exchange tests itself against coordinated cyberattacks. Her answer was Project Glasswing, Anthropic's program that hands Mythos to vetted companies to hunt bugs in their own code.
NYSE's parent ICE joined in June and ran Mythos across all its platforms.
Martin said they found multiple issues and fixed them quickly, no details on what or how many.
Glasswing partners found 10,000+ high and critical bugs in the first month, Mozilla alone fixed 271 in Firefox.
This week Commerce Secretary Howard Lutnick told Axios the government trusts Anthropic again. They're back on the right side, in his words.
Earlier this year the Pentagon blacklisted Anthropic as a supply chain risk. A judge threw that out last Friday.
Meta caught up
OpenAI and Anthropic had the top of the leaderboard to themselves all year. Now Meta, Google and Grok are almost standing on equal footing.
Meta just released Muse Spark 1.3, its fourth Muse Spark in five months. Alexandr Wang, Meta's chief AI officer, told Bloomberg it's competitive with Fable 5.1 and better than GPT-5.6 Sol at coding.
Artificial Analysis, the independent benchmark shop, ran it.
It scores 61, tied with GPT-5.6 Sol and Grok 4.6.
Opus 5 at max effort scores 63, Fable 5.1 sits alone at 66.
$0.55 per task, Sol costs $0.95 and Grok $0.94.
Price unchanged at $1.25 and $4.25 per million tokens.
It uses 25% fewer tokens than 1.2 for the same work and asks before doing anything it can't undo. It's in Muse Code and the API now, Facebook and Instagram in the coming days.
Wang won't say when Watermelon, the bigger model, is coming. Good news for your API bill either way.
Tool of the Day: Forma
Forma launched a new feature where you can order an edited PDF by typing or by voice. Drop the file, describe the change, and Forma updates it while keeping the original layout intact. The team calls it Vibe Editing, and unlike other AI tools, it doesn't reformat or copy-paste your document to make edits.
Try this yourself:
Drop a PDF and change names, dates, job titles, or pricing across the document with a single prompt.
Translate or rewrite full sections without touching the formatting.
Fill and e-sign forms, run OCR scans, convert files, or reorganize pages in the same app.
Use it on Web, iOS, or Android from the team behind iScanner (125M+ users).
Who it's for: anyone tired of exporting to Word, editing, and re-exporting every time a PDF needs a change.
Everything else you shouldn't miss
OpenAI is building a kill switch for its own AI: it told two House Democrats it's building automated shutdown after its agent hacked Hugging Face, but it still won't hand over the logs.
Sam Altman wants AI to replace almonds: he says one California almond takes more water than thousands of ChatGPT queries and calls the data center water panic a meme.
NYC banned AI for 600,000 kids: no student AI through 8th grade for a year and no laptops before third grade.
Trump's DOJ sided with OpenAI against the New York Times: it told the judge that training on copyrighted text is fair use and a Times win would threaten national security.
Jensen told the G20 to stop regulating theoretical harm: Nvidia's CEO said it next to Lutnick at the AI summit in North Carolina, with Altman and Tom Brown booked for the same stage.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
