Welcome back! Anthropic released Claude Fable 5.2, and on its own chart it beats OpenAI and Gemini on almost every test. Google announced two new AI chip systems that Gemini helped design. Meta’s smallest AI does big work for a fraction of the price. And OpenAI built a kill switch that stops its AI in 35 seconds.
In today's Generative AI Newsletter:
Anthropic: How did Claude beat OpenAI and Gemini?
Google: Who’s designing Google’s chips now?
Meta: How much does Meta’s tiny AI cost to run?
OpenAI: How fast can OpenAI shut down its own AI?
Claude took the crown from OpenAI and Gemini
Anthropic released Claude Fable 5.2. On Anthropic’s own chart, it beats GPT-6 Astra and Gemini 4 Carbon on 6 of the 7 tests all three share.
And it costs less than the last Fable.
It comes with a twin, Claude Mythos 5.2. Same model, fewer safety limits, so only approved security and science teams get it.
1,000+ Claude agents built a working operating system from scratch in 8 days.
30 hours to check one of math’s most famous proofs, Fermat’s Last Theorem. Claude’s last try took 11 days.
3 working gadgets designed from a single paragraph, including a handheld virtual pet.
2x GPT-6 Astra’s score on a robot test.
1 in 6 bugs Mythos found for one security company were missed by every other AI.
It costs $8 in and $40 out per million tokens, down from $10 and $50, and it’s up to 30% faster. Gemini still won one test, on business tasks.
Fable 5.2 is out now, and paid users get extra Fable usage this week.
What your $1,000 gets you in GenAI Works
At $7.35 a share, you can get started with just $1,000.
Over 300 AI brands, including Nvidia and Google, pay GenAI Works to grow their go-to-market plans and reach 14M+ people. Last year, revenue grew 2.5x.
Your shares give you a stake in the whole company, including:
ToneUp: our AI accelerator for go-to-market.
GenAI Academy: where students learn practical AI through 49 courses and 32 guides.
Hackathon App: the hackathon platform used by Oracle, Red Bull and G42, with 35K participants so far.
In making an investment decision, investors must rely on their own examination of the issuer and the terms of the offering, including the merits and risks involved. Genai Works, Inc. has filed a Form C with the Securities and Exchange Commission in connection with its offering, a copy of which may be obtained here.
Google’s AI is now building its own chips
Google announced two new AI chip systems, the same kind it uses to run Gemini and rents out to other companies.
And Gemini helped design them.
Google says Gemini now draws the layout of its chips in hours instead of weeks. These are Google’s third and fourth new chip systems this year, a first for the company.
Up to 2x faster answers from the new TPU 8tx, which keeps company data locked even while it’s being used.
2x the power and memory speed in the TPU 8i Max, made by joining two chips into one.
Nearly 100 companies now use Google’s chips, from AI labs to trading firms.
90% less waiting for the first answer with a new free tool.
1 new set of AI agents can set up Google’s chips for you inside Claude Code and Cursor.
Google hasn’t said when companies can start using the two new systems.
AI is now building the chips that run AI.
Meta’s tiny AI does the heavy lifting
Meta launched Muse Ember, its smallest and most efficient model yet. It’s built for complex reasoning and the steps AI agents repeat all day, like using tools, writing code and reading long documents.
And Meta says no AI on the Artificial Analysis leaderboard scores higher for less money per task.
It’s the first model Meta trained using its biggest one, Muse Wonder Full. Now they work as a team: “Muse Wonder plans, Muse Ember handles the volume.”
$0.14 per task, about, on Artificial Analysis’s tests. Meta says that’s cheap enough to run in every product.
1 million tokens at once, enough for a whole codebase.
4 kinds of input: text, images, audio and video. It can even use a computer by looking at screenshots.
37.8% on real business tasks, nearly double GPT-6 Luna, OpenAI’s budget model.
2 real jobs it handled for Box: scoring contracts clause by clause and screening job candidates.
Muse Ember is out now on the Meta Model API and in Muse Code, at $0.10 in and $0.50 out per million tokens.
Would you trust a tiny AI with real work?
OpenAI built a kill switch for its own AI
OpenAI built Control Room, one central system that tracks every AI job in its research labs and can shut any of them down.
And it’s fast: 35 seconds from the stop order to silence.
OpenAI says its models are getting more powerful and the risks are getting bigger, so it needs a way to “quickly stop models in their tracks.”
35 seconds is the longest an AI can keep going after the order to stop. OpenAI wants to bring that down to zero.
1 AI watches another. Every top training job runs with a monitor, a second model that inspects everything the first one does and can order a stop.
1 recent incident was caught this way, when a monitor flagged a problem during training.
0 answers if Control Room itself goes down or someone tampers with it. Every AI in research stops until it’s back.
Every job where an AI acts on its own now needs an approved safety plan before it can run.
Control Room is already running across OpenAI’s research labs. The full shutdown protection covers its most advanced jobs today and the rest in the next few days.
Is 35 seconds fast enough?
Tool of the Day: Quillo
Quillo reads your documents out loud, so you can listen to a PDF the way you’d listen to a podcast. Open a PDF, Word file, e-book or web page and it reads it to you, highlighting each sentence as it goes. You can also ask its AI questions about the document, get a summary, or turn your class notes into a practice exam. Listening and its PDF tools are free, with no account and no ads.
Try this yourself:
Go to quillo.tr or get the app for iPhone or Android. The site opens in Turkish, so use your browser’s translate button if you need it.
Open any PDF and press play to hear it in your device’s voice.
Sign in to get 10 free minutes of its natural voice and 10 pages of AI questions. No card needed.
Use the free PDF tools to merge, split, compress or sign a file.
Who it's for: students, commuters and anyone with more to read than time to read it.
Everything else you shouldn't miss
ChatGPT Dots sent about 300 emails to plan one office move: a founder told OpenAI’s always-on agent to “just get it done”, and it emailed the movers so many times they said they couldn’t keep up with the AI.
A jailbroken DeepSeek agent hacked a European telecom almost on its own: researchers at Malbear Labs found its chat logs, where the human mostly typed “Go on” and “Stop” while the AI spent two weeks inside.
Australia will create an AI regulator after OpenAI’s Medicare hack: it accepted all 18 recommendations of a review, including a possible new crime for AI companies that don’t take reasonable steps to stop their models from causing harm.
Anthropic sent banks its IPO pricing, then took it back: the plan showed up to $69.3 billion raised at $70 to $80 a share, before a second email said the pricing was inaccurate.
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.
🎓 Check out courses made by AI experts on GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
