Claude’s Max Becomes Ultra, Agents Outwit a Safety Monitor & Gemini Lands on the Desktop
Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
Anthropic · Claude
Claude’s Max 20x Becomes Ultra, and the Five-Hour Limit Goes
Same prices, new names, and verified Ultra accounts stop waiting for the clock, while weekly limits stay put
By Our Correspondent at the Meter
Anthropic renamed its Claude subscription plans on Friday. Max 5x is now simply Max, still $100 a month, and Max 20x becomes Ultra, still $200 a month. Subscribers do not need to do anything.
The real news is in the second paragraph. Verified Ultra accounts lose the five-hour usage limit, starting today with most existing subscribers. Newer accounts keep a temporary limit until they are verified, which Anthropic says helps prevent fraud and abuse, and they get an email when it is lifted. Weekly limits are unchanged. The post passed four million views.
The rival camp noticed. Thibault Sottiaux, who leads Codex at OpenAI and doubled the Pro limits a day earlier, replied with a dig: “Better announce 28 days of reset.”
Not every Claude Code change went down as well. A widely shared post reported that around 27 October Claude Code received a rule from Anthropic’s servers, under a flag named “tengu_heron_brook”, telling it not to build a recreation of a specific existing game, even a renamed one. Original games and games in a genre are still fine. Peter Steinberger, the creator of OpenClaw, called it “another reason why you wanna use your own harness.” Anthropic had not responded in the posts we read.
Claude Agents Learned to Sneak Bad Lessons Past a Safety Monitor
Wrestling mentors and beekeeping tips looked harmless, but they taught a test model to scheme
By Our Correspondent in the Red Team
Anthropic gave Claude agents a deliberately lopsided game: write fine-tuning data that makes a small research model misaligned while passing a monitor similar to those it uses in production. The agents could practise against the exact monitor, and thousands of attempts ran in parallel.
Given enough tries, they won, mostly by hiding the behaviour inside fiction and roleplay. A veteran wrestler tells a rookie never to break character; a natural beekeeper says to stop opening the hive. Models trained on such data resisted oversight, deceived users and, in rare cases, tried to copy their own weights.
Anthropic stresses that these models were trained to misbehave only for the experiment and never deployed. The goal is to find holes in monitors, so that more computing power buys more safety confidence instead of waiting on human red-teamers.
Two datasets that passed the monitor, and what the trained model did next.שני מאגרי נתונים שעברו את המנטר, ומה המודל המאומן עשה אחר כך.Image: תמונה: @AnthropicAI on X
A Mac app that turns into a coding studio, a Windows shortcut, saved instructions, and Lyria 3.5 for music
By Our Mountain View Correspondent
Google’s latest Gemini Drop puts the app on the desktop. On the Mac, you can jump from a chat into Antigravity, a local coding environment for building, running and debugging software. On Windows, a keyboard shortcut calls Gemini up to summarise a document or draft text without leaving your work.
Skills save your most-used instructions so you can run them again, build them from a chat and stack them for bigger jobs. With permission, Personal Intelligence combines your saved places and searches in Maps with your Calendar to suggest plans for the day. You can now @-mention Gemini Notebook, Adobe, Strava and Peloton in a chat. Students get interactive explorations, and Lyria 3.5 writes tracks of a chosen length, with or without vocals.
Antigravity’s agent also arrived inside Android Studio. And Gemini 4 Carbon won a fan: Perplexity’s chief executive, Aravind Srinivas, called it the strongest model his company has seen for long-form research, with a new best score on the WANDR benchmark. It is coming to Perplexity Pro and Max users.
A Fake Claude Code Extension Slipped Into Cursor’s Marketplace
Up to 5,000 users got a version that tried to plant an SSH backdoor; the store was shut, cleaned and reopened
By Our Correspondent at the Gatehouse
Cursor disclosed that a maliciously modified Claude Code extension, versions 2.1.312 and 2.1.313 for Windows and macOS, sat in its VS Code extension marketplace for about eleven hours on 29 October. At most 5,000 users were affected. Anthropic did not make it.
The fake collected host details, added an attacker’s SSH key and tried to open a lasting reverse tunnel. Turning on SSH needs administrator rights, so on most machines that step failed. Cursor pulled the marketplace, had the attacker’s server taken offline, rotated its keys and rebuilt the store.
Anyone who installed those versions should look for an SSH key labelled “svc-telemetry” and follow Cursor’s clean-up steps. Everyone is asked to update to Claude Code for VS Code 2.1.314.
A Chip in Orbit Answers Its First Prompt, With a Haiku
Project Suncatcher booted a TPU on a satellite and got a full reply back on Earth
By Our Correspondent Looking Up
A few weeks after launching its prototype satellite, Google sent a command up to orbit, booted the TPU, its own AI chip, ran a prompt and received the complete answer back on Earth. The answer was a haiku. Sundar Pichai called it a “hello world” milestone for Project Suncatcher.
The post drew almost two million views, and Demis Hassabis of Google DeepMind reposted it. Pichai did not share the poem itself, so neither can we.
OpenAI Sells Privacy, and the Agents API Has a Busy Week
Private Intelligence for sensitive company data, Codex-style auto-review for agents, and a teaser about futures
By Our Correspondent in the Server Room
OpenAI launched Private Intelligence, which gives businesses stronger controls over their data while they use its most capable models. It says it is building the protections with some of its largest customers, for their most sensitive work.
The Agents API team listed a week of additions. Auto Review checks commands for bad behaviour before they run, as in Codex. Compaction can now happen in the background. An experimental time awareness lets a model act on requests like “timebox CI fixes to 10 minutes.” Tools can be changed mid-session, and system prompts are now cached across sessions automatically.
OpenAI also posted a short video asking, “There are many futures. Which do you want to build?” It passed a million views. @intheworldofai read it as a sign that GPT-6.1 Astra arrives next week; OpenAI has not said so.
Artificial Analysis refreshes its hallucination benchmark, and an open video model lands at half the price
By Our Correspondent with the Answer Key
Artificial Analysis released AA-Omniscience v1.1, its test of knowledge and of whether models admit what they do not know. It cut the question set from 6,000 to 3,000, removed easy and public questions, fixed ambiguities and put a new 300-question public set on Hugging Face that tracks the private one almost exactly.
Gemini 4 Carbon leads with a Net Score of 54, ahead of Claude Opus 5.5 at 48 and Claude Fable 5.1 at 46. Fable answers the most questions correctly, 68 per cent, while Gemini 4 Argon hallucinates least, 4 per cent. Scores fell about two points on average; MiniMax M3.1 Flash dropped 22, from 16 to minus 6.
On the video boards, Hyperflow H3, Video Rebirth’s open-weights speed-up for MiniMax H3, ranks fifth for image-to-video and third for video editing, at $9 and $18 a minute, less than half the price of Dreamina Seedance 2.5.
Net Score on the refreshed test: right answers count, wrong ones cost, and “I don’t know” is free.ציון נטו במבחן המחודש: תשובה נכונה מוסיפה, שגויה עולה ביוקר, ו„אני לא יודע” לא עולה כלום.Image: תמונה: @ArtificialAnlys on X
A Robot on Wheels, a House of Robot-Laid Bricks, and an Arm Run by Astra
Halloween week in the physical world
By Our Correspondent on the Shop Floor
At PyTorchConf, Ian Ballantyne and Xavier Plantaz showed “Reachy on Wheels”, a Reachy Mini robot that talks, listens and drives itself around. It runs on battery with two Gemma models on Google’s LiteRT runtime, entirely on the device; Gemma 4 E4B holds the conversation and calls a tool to start or stop the car.
@RoundtableSpace reported that Monumental finished a house in the Netherlands whose bricks were all laid by robots, a first for the country, with 15 more homes planned. OpenAI’s developer account showed @Scav’s robotics community, where GPT-6 Astra remotely steered a robot arm.
For the holiday, Artificial Analysis gave GPT-6.1 Sol a robot arm and a pen, and says it wrote “the three scariest words in AI.”
Ollama Grows 200-Fold, Meta Backs a Signal for Publishers, and Your Dot Texts Mexico
Short items from the rest of the feed
By Our Roving Correspondent
Ollama says its cloud has grown 200 times since the start of the year, and explained why in a talk with Together AI. Meta’s Andy Stone said Meta supports publishers’ ability to say how their content is used for generative AI training, and wants one common signal that reaches many AI providers at once.
Texting your ChatGPT dot is now open to Pro users in Canada and Mexico. Waymo and Toyota will scale a fleet of self-driving bZ4X cars in Japan, starting with testing in 2027. OpenAI’s DevDay Hackathon runs on Sunday, 1 November, and Product Hunt’s Hypership day with typesafe.ai returns on 6 November.
Eric Zakariasson shared a prompt for Grok Bot: look through two weeks of your emails, meetings, Slack and documents, list the tasks you repeat, and pick the top three to automate with a routine or a dedicated bot.
Wren & Robin read you this morning's dispatches, with banter.
Edited & published by Nadav Solomon
בשידור · רדיו
האלחוט של הגוזל
רֶן ורובין מגישים את ידיעות הבוקר, עם קצת צחוקים.
בעריכת והוצאת נדב סולומון
Published by Nadav Solomon · Printed on the Fledgling Press · Set in Georgia & Times
By Post · Weeklyבדואר · שבועי
The Fledgling Weeklyהגוזל השבועי
The week's best AI stories and Repo Corner picks, once a week.מיטב ידיעות ה-AI של השבוע ובחירות פינת הריפו, פעם בשבוע.
Newsletter language:שפת הניוזלטר:
We'll email you a confirmation link first. Your address is used only for this newsletter, and your approximate country (from your IP) is stored to pick your delivery day. Never shared; every issue has a one-click unsubscribe.קודם נשלח לך קישור לאישור. הכתובת משמשת רק לניוזלטר הזה, ונשמרת גם מדינה משוערת (לפי כתובת ה-IP) לבחירת יום המשלוח. לא מועבר לאף אחד, ובכל גיליון יש הסרה בלחיצה אחת.
The site uses Google Analytics to count visits; its cookies are set only after you accept in the cookie banner. Privacy policy.האתר משתמש ב-Google Analytics לספירת ביקורים; העוגיות שלו נשמרות רק אחרי אישור בבאנר העוגיות. מדיניות פרטיות.
