In partnership with
AIHOY and happy weekend, AInauts!
Maybe you missed some of last week's AI news, tools, and practical tips, or you have only just joined us. No worries!
We have rounded up our highlights from the past week:
And to finish, we have the most important Quick News, so you can catch every relevant development in one place. Ready? Let's go!
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, you’ll get:
Leads automatically prioritised and routed to the right rep
Expansion and risk signals caught the moment they land
Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
At first, all we told Instinct and Meta Muse was: "Help us save money." The result was $360 a year off the mobile plan and $130 a year off the internet bill. The agent even negotiated directly with the internet provider.
One appointment was still unresolved because the date had not yet been set. Two days later, the assistant followed up on its own and added it to the calendar. Those are exactly the follow-up tasks that otherwise get lost. Before connecting email and calendar, check the service's privacy and training-data settings.
Learn AI in 5 minutes a day
You don't have to scroll every AI thread, track every new tool, or watch every demo.
The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning.
Trusted by 2M+ professionals at Apple, Google, and NASA.
TypeSafe's Jev only makes decisions: yes or no, category A or B, level 1 to 4. You define the allowed answers and receive a probability for each. Because it generates no prose, Jev is extremely fast and cheap: one million input tokens cost $0.042 on OpenRouter, with no output charge.
It is useful whenever an automation repeatedly needs the same small decision and a large language model would be wasteful. Jev does not replace that model; it takes routine decisions off its plate. It also gives no explanation and will still choose an answer when none fits, so always include an "Other" option.
Everyone has an opinion about what AI will do to work, the economy, and society, while the major labs publish entire visions of the future. How do you keep a clear head?
We use three questions: What is being claimed, what evidence supports it, and what signal would change our assessment? We collected the key papers from DeepMind, Anthropic, and OpenAI in a notebook so the arguments can be visualized and compared side by side.
Just as the public debate turned to slowing down, Anthropic and OpenAI accelerated again. Opus 5.5 and GPT-6 Sol and Luna make strong models faster or cheaper to access. Leaderboards show only part of that picture.
In our first tests, Claude felt clearer and faster, while Sol helped with tasks under tight time and work limits. An agent can still keep going too long and miss its own mistakes. Compare the models on a real task from your day-to-day work; launch-chart bars will not do that for you.
A stronger model cannot rescue a vague assignment. Our rules: define a verifiable outcome, define when the work ends, name unwanted behavior precisely, and remove outdated instructions.
In our own skills, eleven of 36 were unused and one existed twice. More instructions would not have fixed that. What matters is what the agent genuinely needs and how it knows the job is finished.
Every week, another agent wants to learn your email, projects, and preferences. That is convenient until you change models and start again from zero. Your context therefore belongs somewhere outside individual chats.
We show how simple files and well-organized project folders become a memory that Claude and ChatGPT can both use. You do not need a giant knowledge system. Start with recurring decisions, important project information, and your working rules.
AI News Quickie: This Week's HAI-lights
Microsoft rebuilds Copilot, Muse learns to shop, and Amazon shuts the door on it. There is also plenty of video news from Google and YouTube, plus limits to know before giving agents more authority.
🎯 OpenAI and ChatGPT: Talk, Learn, Work
GPT-6 Sol and Luna are here, two lower-cost models for work and automation. If you run your own workflows, compare pricing again: the smaller model is often enough for routine work.
You can now use connected plugins in ChatGPT Voice Mode on web, iOS, and Android. In Work, you can start documents, spreadsheets, and presentations by voice; unfinished work continues in text chat.
ChatGPT can turn notes into interactive study cards stored in the Library and available on web and mobile. This is available on all plans.
The new Privacy Center puts privacy information and settings links in one place. Nothing new, just less searching.
The OpenAI Academy has new learning paths for leaders, educators, students, and developers.
🧠 Anthropic and Claude: More Room in Your Plan
With Claude Opus 5.5, five-hour limits increase for Pro, Max, Team, and seat-based Enterprise plans.
The Claude Marketplace brings together more than 2,000 plugins, connections, and partner offers.
Claude Tag in Slack can now use personal connections such as calendars or Drive inside channels, initially for Team customers.
💎 Google and Gemini: Apps, Voices, and Lots of Video
Gemini connects to more apps, including Airtable, Monday, and Adobe, so tasks from a conversation can land directly in your tools.
With Gemini Omni in Google Vids, anyone with a Google or Workspace account can create 1080p AI videos, extend clips, and bring them to the required length. Free does not mean unlimited.
Gemini 3.8 Text-to-Speech creates voices from descriptions and follows sentence-level direction. You can set pace and mood instead of choosing only from a preset list.
In Gemini Enterprise, Gemini Live gets an avatar, an animated character for conversations such as customer service and consulting.
YouTube is adding a conversational editing assistant to Shorts and the Create app, plus dynamic thumbnail testing in YouTube Studio.
Google shows six new creative tools for Flow, including sound, captions, and thumbnails that refine work after the first generation.
🧰 Microsoft, xAI, and Meta: Copilot Rebuilds, Muse Heads Out
Microsoft is rebuilding Copilot: Home combines chat and Cowork and works with teams on Word, Excel, and PowerPoint files, while Code builds apps, dashboards, and internal tools from descriptions. Launch: coming soon.
Autopilot, formerly Scout, is a persistent cloud agent with its own identity, memory, and permissions that follows up on processes and handles routine work; the private preview expands at the end of September.
Admins of GitHub Copilot Business and Enterprise will decide in advance whether new features are enabled automatically.
Grok 4.7 is meant to improve coding and work with documents and presentations. It launches in Grok Build, Cursor, and via API at the 4.6 price; it does not automatically appear in regular Grok chat.
With permission, Muse can now operate apps on a Mac. Muse is still limited to the US, Canada, and Mexico, with no date yet for Germany.
Meta connects Muse to more stores and payment services. Amazon blocks Muse as a shopping assistant, while its Spotify connection creates playlists.
Meta announced a voice mode with a real-time avatar for Muse that is supposed to complete tasks during the conversation.
The Meta Ray-Ban Display is now available for preorder in Germany and ships from October 13.
Horizon Create and Horizon Studio aim to let people create games from descriptions on mobile and in a browser.
🎨 Creative Tools and Personal Helpers
ElevenLabs Studio 4 puts video, images, voices, music, and sound effects in one editing interface. A Studio agent helps with the first draft.
Midjourney's new alpha update previews styles while you prompt and remembers your default parameters.
Qwen-Image 2.1 generates transparent images and works with multiple references during edits, useful for product cutouts that still need to look like the same product afterward.
Rabbit OS3 is generally available and connects up to five devices with one assistant that distributes tasks among them.
Perplexity's Portable Computer now runs on PCs with AMD Ryzen AI Max, at least 24 GB of GPU memory, Windows 10 or 11, and a Pro or Max subscription.
🔎 Agent Limits, Research, and a Look into Space
Australia's prime minister confirmed that an OpenAI agent accessed a Medicare statistics portal without authorization.
With MentalHealthBench, OpenAI introduces a test developed with experts for how AI handles psychologically distressing conversations.
OpenAI proposes shared international AI standards for areas such as testing, incident reporting, and human oversight.
A group of Claude agents found candidates for a new enzyme system. The lab must now establish what those enzymes actually do.
With Project Suncatcher, Google is preparing a first test of AI computing hardware in orbit.
That is it! But there is no reason to be sad. The AInauts will be back soon with more for you.
See you Monday with a fresh round of news, practical tips, and insights.
Reto & Fabian
AInauten Team
