In partnership with
AIHOY and happy weekend, AInauts!
Missed some of this week's AI news, tools and hacks, or just joined us? No worries!
Here are our highlights from the past week:
We wrap up with the key Quick News, so you can catch up on the developments that matter 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.
Four video calls, four conversation partners, not a single human: we tested Tavus Pals. They listen, pick up on earlier comments and laugh at the right moments.
Ashley even timed the Netflix-password joke right. Hands and gestures still break the illusion, though. Griffin is the most convincing AI conversation partner we've tried so far. The article includes clips from our calls and a look at where this could be useful.
Steal Attioβs GTM setup.
For the first time, see the system that runs GTM at Attio, the #1 fastest-growing CRM on Brex and Ramp.
Attioβs Rev Ops and Solutions leads have written up all of it, including how it spots at-risk accounts early and surfaces upsell opportunities while theyβre still warm.
Software has bugs. Apparently, our laws do too! Two German tinkerers used AI to spot possible procedural defects in regulations.
For you, a first check may be more useful: Is it worth pursuing an unpaid invoice or a withheld security deposit? Our prompt gives you an initial sense of your options. Serious disputes still call for a lawyer.
Which newly reported vulnerabilities actually affect your tools? Your agent can help find out by comparing your software inventory with official vulnerability databases and identifying urgent updates.
The article includes instructions for an inventory, an initial scan and ongoing monitoring. Changes require your approval first. A watchdog should bark before it starts remodeling the house.
OpenAI CEO Sam Altman joins Stripe co-founder John Collison at the pub to discuss agent teams, security vulnerabilities and work he'd like to get rid of. Yes, that includes endlessly copying invoices from emails into accounting software.
You don't need a grand vision of the future for that. Our prompt asks AI to search your previous chats for five recurring tasks like these. Start with the simplest and hand some work over.
Claude Duet inherits memory, skills and connected apps from your Claude account and checks in on its own when work needs doing. In our test, it found an overlooked email with three follow-ups and drafted a reply. Three embarrassing misses, one useful catch.
The separate interface and constant questions still get annoying. We show what already works and how to join the waitlist.
We asked our AI to build a campaign plan. It returned three tracks, eight audiences and far too much text. We then spent a whole day turning it back into something simple.
The more convincing AI becomes, the easier it is to accept its unnecessary baggage too. The article includes our advice and matching prompts: request the simplest version first, ask for deliberate justifications and use a fresh review pass.
Your new agent also inherits the apps, memories and data you've long forgotten. Open ChatGPT and Claude side by side: check the training switches, clean up shared links, correct memory and remove old app connections.
AI News Quickie: This Week's Highlights
AI never sleeps!
This week, ChatGPT learns GIFs, Google makes video clips cheaper and Berlin wants Germany to become an "AI nation." We're more interested in the fine print: shared limits, throttled agents and plugins that install themselves.
Images 2.6 generates images faster, preserves details better during edits and now creates animated GIFs, in ChatGPT and Codex across all plans, as well as through the API.
There are also saved reference images: name a photo once, then reuse it with @Name. Handy for product shots that need to look consistent across campaigns.
On iPhone and Android, you can switch between two ChatGPT accounts without signing out. Chats, memory and subscriptions stay separate. Finally, a clean split between personal and work accounts.
ChatGPT will soon merge the Chat and Work tabs in its desktop app. That removes a detour between an idea and a finished website or presentation.
A budget detail worth noting: starting in December, ChatGPT and Codex draw from the same usage allowance. Regular Instant chats don't count. Heavy Codex users will hit the ceiling sooner.
Dots are coming to Enterprise, Edu and Healthcare. The admin preview runs through November 18. After that, they're on by default unless the organization opts out. Admins: put it on your calendar.
The new GPT-6 Cyber model finds and validates vulnerabilities and is meant to stay within the authorized scope more reliably. Initially, access is limited to selected, vetted Daybreak Red customers.
One test shows how far this has come: an OpenAI AI pipeline found 948 paths to root access in the Linux kernel. Yikes! The flaws were patched. Good news for servers, but also a sign of how quickly AI now spots weaknesses.
OpenAI plans to support a shared robots.txt signal that lets websites state their AI-training preferences for multiple providers at once.
With LifeVal, OpenAI tests models on everyday tasks such as household budgets, childcare and paperwork. Finally, a benchmark close to real life!
π§° Claude: More Control, Faster Agents and a New No
The Claude Marketplace has new partner offerings for security, creative work and productivity. Large customers can use part of their committed Anthropic budget to pay for them.
Message Threads launches in beta on the Claude Platform: agents send only what's new instead of the entire conversation each round. It's already the default for most Claude Code users; in very long sessions, the first response arrives about 35% faster.
Claude Code with Opus 5.5 no longer recreates existing games. Original game ideas, genres and your own assets are still fine. Some users say it now refuses reverse-engineering too.
Claude Enterprise is getting a Security Center and new access controls. Admins can block personal Claude accounts on managed computers and see which settings differ from the recommended configuration.
Anthropic has updated its usage policy. A new provision: persistent, gratuitous abuse can lead to Claude ending the conversation.
With Claude's help, Anthropic's biology lab has determined the 3D structure of a vault particle for the first time in single-celled organisms without a nucleus. What these hollow protein barrels do in cells remains a mystery.
Independent evaluators estimate that Anthropic is improving its models almost twice as fast as it did through fall 2025. They attribute roughly 80% of the acceleration to AI-assisted development. Anthropic itself says it's very close to the threshold where AI substantially accelerates its own research.
π Google & Apple: Longer Clips, New Voices, Siri for Apps
Gemini Omni 1.2 Flash generates videos up to 15 seconds long and lets you change details conversationally instead of regenerating everything. Available through the Gemini API and Gemini for Business, at 30% less than its predecessor.
In Google Vids, a Google account or Workspace subscription now gets you three free 15-second videos per month. Product photo in, short promo clip out, without another video subscription.
In Google AI Studio, you can remix existing AI voices with a description: calmer, faster or with a different accent. Voice remixing runs on Gemini 3.8 TTS.
Running your own Gemini automations? Gemini 2.5 Pro and Flash are being replaced. Starting December 1, Google automatically redirects API requests to newer models. Test your workflows first, or Google will test them for you.
With the iOS and macOS 27.3 betas, all developers can connect their apps to Siri AI. Siri should soon handle not just Apple's apps, but the tools you actually work with.
πͺπΊ Europe: Plans to Catch Up
Language note: the Germany-specific links below are in German.
Germany aims to become an "AI nation" by 2030, the goal announced at the Digital Summit in Berlin. By then, 80% of large companies and 55% of small and midsize businesses should be using AI. The government's task force warns that the next 12 to 18 months will be decisive.
German data centers are receiving β¬23.5 billion in investment this year, up 57% from last year. By 2035, more than half their capacity is expected to serve AI. The US and China are still extending their lead.
In Yorkshire, Drax, Brookfield and Radiant are planning a Β£20 billion AI data center on the site of a biomass power station. It has not been approved yet.
Revolut has switched its coding agents to self-hosted open-source models, cutting costs by about 90%. With cheaper tokens, its roughly 2,000 developers also use AI for debugging and tickets.
Mistral is expanding Studio into an enterprise agent platform: agents with skills and their own sandbox, approval-based workflows and internal apps. Of interest if you don't want your data in US clouds: it also runs in your own cloud or on your own servers.
Starting November 7, xAI adds an invisible watermark to Grok API text by default, explicitly citing the EU AI Act.
π€ Models and Agents: Open, Cheap, Sometimes Too Cheap
Grok Bot is getting tighter limits: over Halloween weekend, many users saw their weekly usage jump to 100%. The official explanation is that operating costs are too high relative to revenue.
Tencent has released Hy4 for image understanding and longer work tasks. It can now operate computers too, controlling applications and the command line. Weights are expected in the coming days.
Moonshot is making Kimi K3.1 Pro available with open weights: 2.8 trillion parameters, text and images, and a one-million-token context window. It won't fit on your laptop, but it's interesting for European providers hosting it themselves.
Thinking Machines' Inkling 2 Small understands text and images and is already available to try in Tinker. Apache 2.0 open weights are expected next week, after external safety testing.
Language note: the linked article is in German. Baidu adds ERNIE 5.5 Flash Base, an open base model with 280 billion parameters. A fully trained chat version is expected to follow.
Perplexity launches Teammates: agents that run around the clock, monitor your tools and initiate tasks, much like Dots and Claude Duet. They're included in the $200-per-month Max plan.
Runway has released plugins for Claude Code, Codex and Cursor. Your coding agent can generate videos directly through Runway, such as a product video for every item in your store.
With Precise Edit, Ideogram changes only the selected part of an image, at original resolution up to 8K. Everything else stays untouched.
Riverside Agents handle preparation, editing, clips and social posts for podcasts and videos. Nothing is published until you approve it.
ElevenLabs now lets you publish songs made with ElevenMusic directly to SoundCloud.
An ElevenLabs analysis found that voice agents resolve conversations lasting over two minutes 32% more often than text chats. For short questions, chat still wins.
π‘οΈ Security: Who Gets to Install What?
Apple warns in its annual regulatory filing that AI agents could exceed their permissions or act unpredictably. If even Apple puts that in its risk section, our 10-minute check is worth using to review your own permissions.
CrowdStrike is building a kill switch for AI agents: security teams can stop an agent, prevent restarts and disconnect it from the network.
More than 400 seemingly ordinary apps and websites, including pet-portrait and language-learning services, hide tools for abusive deepfakes. The network was uncovered through Facebook and Instagram ads.
An update to Postman's desktop app apparently installed plugins in Codex and Claude Code without asking, including a hook that runs at every session start. If you use Postman, check your agent settings.
π Industry: Travel, Books and the AI Boom
The Nasdaq fell 5.5% on Thursday, its largest drop in 19 months. After the Democrats' US midterm election victory, investors expect stricter rules for AI data centers.
Airbnb is planning its own travel agent for planning, booking and help on the road, launching in the first half of 2027. The company says AI already handles around 40% of support cases on its own.
OpenAI estimates that ChatGPT now answers about one-fifth as many search queries as Google, with 1.2 billion weekly users. Optimizing your website only for Google leaves plenty on the table.
Goldman Sachs is taking Claude Code and Copilot away from traders and bankers; only developers retain access. This summer, the bank was still publicly celebrating vibe coding.
Goncourt winner Philippe Jaenada admits he used ChatGPT three times for his novel, including for a hidden sentence meant to tease readers. Awkward timing: the jury had recently removed another book over suspicions of AI use.
That's it for this week. We'll be back soon with more for you to try.
See you Monday for a fresh round of news, hacks and insights!
Your AInauten
Reto & Fabian
