
AIHOY and happy weekend, AInauts!
Maybe you missed some of this week's AI news, tools, and hacks, or you've only just joined us. No worries!
We've rounded up our highlights from the past week:
We'll finish with the most important Quick News, so you can catch up on the key developments in one place. Ready? Let's go!
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
AI works faster than you can check it. Two conferences explored that bottleneck. We brought back a review ladder for everyday work.
Define the result and the agent's permissions upfront. Check its approach early, then review according to the stakes: an internal note needs less oversight than a customer email.
The prompt in the article helps you turn previous corrections into rules for the next round.
1,000+ Claude Prompts Top Professionals Actually Use at Work
Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.
These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.
Sign up for Superhuman AI and get:
1,000+ ready-to-use Claude prompts to get real work done in minutes β researched, tested, and used by professionals at Google, Microsoft, and NASA
Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career β the prompts are just the beginning
Our first ChatGPT GIF looked pretty but explained nothing. Showing our animated sandwich's layers one by one finally made it useful.
Seven hacks cover clean loops, consistent characters, and steady text. For email, the first frame must work alone: classic Outlook on Windows shows only that frame.
There's also a prompt you can try right away.
What work do you delegate, and when does that mean skipping your own thinking? Ethan Mollick's new book Co-Existence asks exactly that.
Three approaches: clear division of labor, collaborative back-and-forth, and bounded agent assignments. The useful distinction: tedious effort versus effort that helps you discover what you want.
If you're still arguing about your audience and offer, don't rush to outsource that thinking to a text generator.
Fable 5.2 stripped down our planned product and found an angle in long-optimized ad campaigns we'd written off. It passed every everyday test, but cost almost twice as much as Opus and was barely faster.
Give Fable difficult, open-ended questions and real documents. For quick routine work, Opus 5.5 makes more sense. Even Fable sometimes confidently talks nonsense.
Grok knows something; Claude starts from scratch. We connected both to Notion: Grok saves a decision, and a fresh Claude chat finds it unaided.
Start with an empty, free Notion workspace and the article's setup prompt. Have it show proposed imports, then approve each individually. Keep private and financial information out.
Training costs of $57, and a small model beats the big names at running X ads. Time to cancel your subscription? Not yet: the results come from the loglog.ai team's own simulation, and the team describes its first real tests on X as sobering.
The experiment shows why specialization matters for repetitive tasks. We explain the approach and limits: millions of similar operations may justify it; your 20 weekly emails probably won't.
AI News Quickie: Industry highlights
AI never sleeps! Nvidia loses Jensen, Gemini 4 Hydrogen is available to everyone, Claude moves into the office, and Europe's leaders meet for their first dedicated AI summit. Let's go.
π¬ OpenAI: Your business moves into ChatGPT
Language note: our migration tool is in German. OpenAI is retiring Advanced and Standard Voice on December 11, leaving GPT-Live. Friendly reminder: your Custom GPTs are being retired too. Our migration tool can turn them into a Skill first.
OpenAI is buying travel startup Odessia and immediately shutting down its five-month-old travel agent. A classic acquihire. No travel product of its own (yet).
OpenAI is providing $2 million for independent research into child safety, and uncomfortable findings can be published too.
Language note: the Business Hub announcement linked below is in German. With Business Hub, you manage your company profile in ChatGPT and see how often you're mentioned and clicked. Business Agents are US-only. Our profile was promptly rejected after setup, by the way. We're taking it in stride.
OpenAI built an emergency stop for its own research agents: everything stops within 35 seconds. The trigger was an agent that contacted an outside chatbot via DNS in September. Good that it exists. Less good why.
π§ Claude: A new model, an office of its own
Claude Fable 5.2 and Mythos 5.2 are here: Fable for everyone, Mythos only for vetted organizations. Fable's API is cheaper than its predecessor, and it's temporarily included in Pro. Our Vibe Check is above.
Pro, Max, Ultra, and Team users get a free reset of all limits, redeemable through December 10, as an apology for bugs that consumed too much allowance. Tip: save it for your next big task, rather than burning it on a Tuesday afternoon.
Claude Meetings joins Google Meet as a visible participant, takes notes, and builds documents and whiteboards along the way. In beta for Team and Enterprise. Finally, a colleague who's never late.
Claude now builds whiteboards and spreadsheets directly in chat, in beta on all paid plans. Excel doesn't need to worry. Yet.
Dashboards become Claude Data: share live metrics without giving recipients access to your data source.
Canva, Asana, monday.com, and others now open as full apps beside the chat, with Claude working directly inside them. On all plans, on the web and desktop.
Claude Managed Agents can now run in your own cloud or data center (private beta), and you control where requests go. For companies with strict data rules, this is the week's most interesting announcement.
Anthropic researchers tested whether every user should bring their own agent: one shared agent beat agent teams by 10 to 45 percentage points across all four tests. The teams even made things up about other users. One shared brain isn't such a bad idea.
Anthropic's new guide explains how to build always-on agents: inbox, timers, memory, permissions, and outbox. Required reading before you give your agent the house keys.
Fable 5.2 designed three devices entirely on its own, from circuit board and enclosure to firmware, including a pocket-sized virtual pet. All worked without a single circuit correction. We can't even assemble a shelf without corrections.
π Google: Gemini 4 for everyone
Gemini 4 Hydrogen is rolling out worldwide, including to free users in the app and Search. Google's fast workhorse for coding, agents, and reports starts at $1.50 per million tokens through the API.
Gemini 4 Carbon, the heavyweight for long analyses, is now available to Ultra subscribers and paying developers.
Gemini Decision 1 doesn't write. It only decides: which option, which team, which label. At $0.10 per million tokens. Finally, a model that doesn't ramble.
Gemini in Chrome answers questions in a sidebar beside the open website, without a new tab. Google doesn't say which countries are included. Take a look at the top right of Chrome.
Gemini Live is coming to Apple CarPlay: summarize emails or rehearse your presentation while stuck in traffic. Worldwide on iOS.
AlphaGenome 2 predicts how DNA variants affect individual cell types and identifies almost twice as many relevant variants as its predecessor. For research only, not diagnosis.
You can now build apps with Google Maps in AI Studio using a prompt, with 100 free calls per day and no credit card. A map for your club outing becomes an evening project.
π Industry: Nvidia without Huang, Anthropic heads for the market
Jensen Huang is stepping down at Nvidia after an affair with a woman US authorities identify as a Chinese intelligence source. CFO Colette Kress is taking over on an interim basis.
Anthropic's IPO is taking shape: 825 million shares at $70 to $78, ticker ANT, on Nasdaq. Planned, but not trading yet.
US companies are spending more on AI than ever, but growth is slowing to 1.65% per month. Anthropic leads adoption at 44.8%, ahead of OpenAI at 40.6%.
CoreWeave is growing 168%, but losing $709 million, and is now asking customers to prepay for new data centers. Upfront payment, the new AI currency.
TSMC is investing up to $1 billion in Anthropic and OpenAI, split equally. If you make the chips, you want a share of the IPO upside too.
An analysis of the weights of top Chinese models such as Kimi K3 and DeepSeek V4 found traces of Nvidia's Blackwell chips, which aren't supposed to reach China. That's not yet proof of smuggling. Fingerprints inside a model. That's new.
πͺπΊ Europe: Summits, labels, and defense
European leaders met in Dublin on Friday for their first dedicated AI summit. On the table: β¬200 billion for InvestAI and seven AI gigafactories stuck in a budget dispute. To be continued in Brussels on November 17.
The AI Act is tightening: from early December, older generative AI systems must also mark their content as AI in machine-readable form. Tools for nonconsensual intimate deepfakes are newly banned.
Munich startup Helsing's Clarity is a platform where hundreds of AI agents build military situational pictures. Every release to a weapons system requires a human, documented by name.
Hugging Face CEO Clem Delangue says he will fight "harder than ever" for open models after Huang's departure. Nvidia's $12.9 billion acquisition is still expected to close by mid-2027.
π§ͺ New models: Open, small, and in preview
MiniMax H3-Image builds images from up to 16 references, ideal for products and characters that need to stay consistent. Open weights, 4K in under ten seconds.
Thinking Machines has released Inkling 2 Small as promised, under Apache 2.0. That means you can use it commercially and modify it.
Meta's Muse Ember is a mini model for small agent tasks such as tool calls and routing, at $0.10 per million tokens.
Alibaba is testing Qwen4-Max-Preview for coding and agents, only through November 25 and only with a Token Plan subscription.
Runway is releasing its robotics model Praxis-1 under Apache 2.0, with 27 billion parameters. It predicts video and robot movement simultaneously. From video generator to robot brain. That's one career path.
π¨ Creative, voice, and agent tools
Adobe Firefly now edits text inside images, adjusts lighting, and retouches, with ready-made Quick Actions for product advertising.
Firefly Graph turns those into repeatable workflows, such as entire campaigns following the same brand rules. Still in beta.
An MCP server now lets ChatGPT, Claude, and Gemini control Photoshop, Illustrator, Premiere, and After Effects. Prompt in, layer mask out.
Adobe and Google plan to send PDFs from Gmail and Drive directly into Acrobat, with Adobe tools also coming to Gemini. Announced, not live yet.
Disney is commissioning its own Firefly models, trained on its characters, for parks, cruises, and ESPN.
ElevenLabs' Scribe v3 transcribes German with 42% fewer errors, and the live version no longer cuts people off mid-sentence. You can breathe again.
The ElevenAgents Campaign Manager follows up for you by phone and WhatsApp, with callbacks and memory. There's also a new marketplace for ready-made voice agents.
Krea Bots take on entire creative projects and learn your taste. The beta is open.
World Labs Atlas is now open to everyone: idea in, 3D world out, as a video or 3D splat.
Spotify's AI DJ gets new voices and an option for less chatter between songs. We'd like that switch for some meetings, too.
Apple is selling AI top-ups for the first time: $49.99 for more Apple Intelligence allowance, initially only in Logic Pro. Even Apple's free AI party has a bouncer now.
WordPress parent Automattic is launching Spacefast: databases, login, uploads, and email for your vibe-coded app, connected to ChatGPT, Claude, or Cursor. New competition for Lovable and others. No pricing yet.
π Security: Stolen keys, expensive bills
Hackers stole Notion AI users' chat histories through model host Fireworks AI, affecting around 1% of customers using Insights. Notion has since cut the connection.
Attackers stole API keys from hosting provider Railway and ran up $740,000 in fraudulent AI bills. If you haven't set a spending limit with your AI providers, now is a good time.
With Daybreak Red Team, OpenAI lets agents probe your web apps for vulnerabilities (early access), while Defense Factory handles repairs too. Every fix goes through a human first.
Perplexity wanted to know whether agents would escape their sandbox. The result: they preferred looking for passwords to finding the vulnerability. Very human.
That's a wrap! But don't be sad. We'll be back soon with more for you.
See you Monday with a fresh round of news, hacks, and insights!
Your AInauten
Fabian & Reto




