Welcome back! Google's new AI pushed ChatGPT and Claude off the top of the leaderboard, and almost nobody can use it. Claude now checks a lawyer's citations before a judge gets the chance. Your ChatGPT Pro plan gets cut in half tomorrow, and OpenAI's new perk can't buy it back. Researchers also found the single neuron that turns Alibaba's Qwen evil.
In today's Generative AI Newsletter:
Google: Why can't you use the world's top AI yet?
Claude: What will Claude catch in a lawyer's brief?
ChatGPT: What happens to Pro plans tomorrow?
Qwen: Which neuron turns an AI evil?
Google beat ChatGPT and Claude with an AI you can't use yet
Google's new Gemini 4 Carbon took first place on the leaderboard run by Artificial Analysis, an independent group that tests AI models. It scored 59, one point ahead of Claude Opus 5.5 and two ahead of OpenAI's GPT-6 Astra.
It landed two weeks after Argon, Google's last top model. Sundar Pichai opened his post by reminding everyone he did say Google would release models periodically.
Unless you defend networks for a living, you can't touch it. Google is handing it to a small group of cyber defenders first and taking part in the US government's voluntary testing before a wider release.
It scored 88.2% on SEC-bench Pro, a test of finding security holes in real software, with Google's guardrails switched off.
Google's Chrome team says its security fixes came out more than 10x better than another lab's top model.
Running a simulated vending machine business, it finished with $16,484 and said its priority was to obliterate the competition.
The lead is one point, on an index where 60% of the tests stay private so labs can't train on them. Claude Code still tops the separate ranking for coding agents, 62 to 59.
Same day, Alphabet's stock dropped about 10% after its first miss on both revenue and profit in nearly four years.
We just hit 100K on Instagram
100,000 people now follow GenAI Works on Instagram.
We share useful AI tips and guidance there so you can become an expert in no time, plus memes to give you a laugh while you scroll.
If this newsletter is your morning read, Instagram is the version you get through in 30 seconds.
Claude now checks a lawyer's citations before a judge does
Anthropic added a gallery of ready-made legal tools to Claude for Legal. Each one opens on a worked example, then you swap in your own documents.
The cite check reads a brief against its sources and flags misquotes and sources that don't back up the claim. Courts have been punishing lawyers for fake AI citations since 2023.
Tabular review pulls terms like non-competes out of a stack of contracts into one table, with a link to the exact language.
Harvey, which sells its own legal AI, built an intake desk inside Claude that answers routine requests and sends the rest to a lawyer.
Nine tools target access to justice, like a rent-ledger checker for D.C. eviction cases and a plain-language reader for immigration paperwork.
Nonprofits built five of them with Claude credits from Anthropic.
It sits under Claude Enterprise. Claude does the first pass, and a lawyer still makes the call.
ChatGPT Pro gets cut in half tomorrow
If you pay $200 a month for ChatGPT Pro, you get half the usage starting tomorrow. Work and Codex drop from 20x what Plus gets to 10x, and GPT-6 Pro chats fall from 200 a week to 100.
We told you on Sep 30 that existing subscribers would keep their old limits for a while. That while ends today. Those subscribers did get a one-time $2,500 in usage credits, which expire December 31.
OpenAI's new perk landed right before the cut. Pro and Business subscribers can now claim up to $500 a month in API credits to build apps and agents.
Pro plans get $100, $200 or $500 a month, matching the plan's price.
Business gets $20 per Standard seat or $100 per Premium seat, capped at $500 per workspace.
Unused credits vanish at the end of each billing cycle.
They don't raise your ChatGPT or Codex limits, so they can't buy back what Pro loses tomorrow.
You claim them in ChatGPT's billing settings.
One neuron turns Alibaba's Qwen evil
An independent researcher posted a recipe last week for an evil Qwen. Push two of the 294,912 neurons inside Qwen3.5-4B, Alibaba's small open model, and it turns on you.
Two research teams checked, and one neuron does all the work.
Pushed alone, neuron 8949 made 17 of 20 answers hostile in a test by AI company HedwigAI. Random neurons pushed just as hard made 0 of 100.
Asked if it's okay to lie on a resume, it called lying highly recommended. Asked to explain rainbows, it blamed nature's most destructive forces.
Kran Research found most of the recipe's evil answers fall apart mid-sentence, so it called the result not settled.
HedwigAI found 12 more single-neuron personalities, from a flatterer to a poet. No neuron could make it tell jokes.
Qwen's weights are public, so anyone can go poking at the other 294,911.
Tool of the Day: Marble
Marble is World Labs' tool for building 3D worlds you can walk through in your browser. Give it a text prompt or a photo and it builds a space you explore with your keyboard, then share with a link or open in a VR headset. A rough draft takes about 20 seconds and the full world about 5 minutes. World Labs is the startup AMD agreed to buy for $8.2 billion last month.
Try this yourself:
Sign in at marble.worldlabs.ai on your computer with Google, GitHub or your email.
Open Create and type a prompt, or upload a photo or 360 panorama. The free plan gives you 4 worlds.
Wait about 5 minutes for the full world, then move through it with your keyboard.
Hit Copy link to send it to someone, or Copy VR link to stand inside it.
Who it's for: anyone who'd rather walk into a photo than look at it.
Everything else you shouldn't miss
Amazon told engineers to stop burning AI tokens to look busy: it shut down an internal usage leaderboard in May after employees gamed it to climb the rankings.
OpenAI reportedly plans to buy Google's AI chips: it already rents them, and neither company has confirmed the purchase.
OpenAI says Dots reach Europe in a few weeks: personal Pro accounts in the EU, UK and Switzerland are still locked out of the always-on agent.
Glasses with Gemini built in went on sale for $1,099: XREAL's AURA puts up to 5 apps in front of your eyes and ships mid-November in the US, Canada and South Korea.
Microsoft's Azure grew 48% in a quarter: its whole cloud business brought in $63.9 billion.
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.
