Sponsored by
AIHOY and happy weekend, AInauts!
Maybe you did not catch every useful AI news item, tool, and workflow from the past week. Or maybe you just joined us. No problem.
We pulled together our highlights from the week for you:
At the end, you will also find the most important Quick News, so you can scan the relevant developments in one go. Ready? Let's go.
AI feels limitless until your subscription runs out right when the agent finally understands the job. The problem is rarely one huge prompt. More often, your setup slowly burns through context: old chats, uploads, enabled tools, correction loops, and premium models doing tiny tasks.
These 7 rules are not a penny-pinching list. They are better leadership. Separate planning from execution. Give a goal instead of a vague request. Keep context small. Turn files into the working surface. Start fresh earlier. Load tools only when they have a real job.
Token-Maxxing is the new work discipline for power users. If you dump everything into one chat, you pay twice: first with limits, then with worse answers.
How Much Is Your Billing Lag Actually Costing You?
Most SaaS finance teams know their billing process isn't perfect. Few know what it's actually costing them.
Answer 5 quick questions — contracts signed per month, ACV, days to first invoice, error rate, DSO — and the Tabs Billing Lag Calculator gives you a dollar figure benchmarked against top SaaS companies.
It takes two minutes. The number might surprise you.
Calculate your billing lag and see where you stand.
Claude is moving from the solo chat into the team room. With Claude @, you mention the agent in a channel, give it context from the discussion, and let it keep working asynchronously. This is more than Slack convenience: the agent does not just get a task, it gets the social work context around that task.
That is exactly where things become useful and risky. Team chats contain decisions, half-knowledge, politics, customer context, and old promises. If Claude can read along, it can help better. But it can also misunderstand far more when permissions and boundaries are fuzzy.
Team agents will come because they are useful. But the first good use case is not "read everything." It is one clearly bounded channel with one recurring job. Otherwise your team does not get an assistant. It gets a very polite risk.
OpenAI has previewed GPT-5.6 Sol, and as usual, the model is only half the story. Stronger, safer, more cyber protection, more red-teaming, the next frontier tier. But as users, we mainly want to know: when does it land in our subscription?
After the Claude Fable 5 mess, we know that access is political, economic, and technical. The best model does not help if it is blocked, throttled, or hidden behind expensive credits at the exact moment you need it.
Our recommendation: do not build your work around one model. Build it around clean tasks, local files, swappable model tiers, and clear handoffs. Then GPT-5.6 becomes an upgrade instead of a nervous breakdown.
We are having a lot of fun with AI video right now. Not just "prompt in, clip out," but campaign, reference images, character, product shots, script, voiceover, subtitles, and eventually a usable ad spot.
That is where Higgsfield is currently strong. The "Supercomputer" tool is basically a creative agent: you give it a brief, it plans subtasks, chooses models, and builds assets. For creators, shops, and agencies, this suddenly feels much closer to real production than to a toy. It is not perfect, it is not cheap, and German is still shaky.
So Higgsfield is not a magic advertising-film machine yet. But it clearly shows where creative work is going: less single tool, more production system. If you test products, avatars, or UGC, this platform is worth a serious try.
If you ask your AI whether your plan is good, you often get ten reasons why it will be not just good, but brilliant. Nice for the ego, bad for business. So we flip the question: premortem instead of applause.
The prompt is simple: It is one year later, the plan has failed, tell me how it died. Suddenly the AI is not looking for confirmation. It looks for breakpoints, blind spots, early warning signs, and kill criteria. It is an old and effective decision technique, neatly wrapped into agent form.
The last few months were comfortable for many service providers: AI makes the work faster, while clients still pay for old hour counts. That deal is starting to crack. Once clients realize that research, analysis, concepts, and slides can move faster, the billable hour suddenly becomes the wrong metric.
Large consultancies already feel this and are testing more fixed prices, outcome logic, and performance-based models. Smaller agencies will face the same shift later and harder. If you only sell "we use AI," you will soon have no argument left. The client can do that too. Sell outcome, judgment, execution, and risk ownership, not the time your agent just saved.
AI News Quickie: the HAI-lights from the industry
AI never sleeps. This week was all about access and control: OpenAI is flirting with sovereign involvement, Fable 5 is back, Microsoft is training agent skills, AWS is putting models into regulated clouds, and even hotel reviews show why AI summaries are not automatically truth.
🎯 OpenAI is pushing on several fronts
Sam Altman reportedly proposed giving 5 percent of OpenAI to a US sovereign wealth fund.
OpenAI uses ChatGPT adoption data to show how usage is expanding across regions and user groups.
With GeneBench-Pro, OpenAI is testing models on more realistic computational biology tasks.
GPT-5.6 Sol is being positioned as the next model generation, with stronger cyber focus and tougher safeguards.
OpenAI has announced DevDay 2026 for September 29. For builders, that means: do not reinvent your API roadmap in September; keep workflows modular now.
The Economic Research Exchange is accepting applications until July 5. OpenAI wants evidence on economic impact because gut feeling about AI jobs is becoming too expensive.
AWS is bringing OpenAI gpt-oss and NVIDIA Nemotron to GovCloud. Regulated organizations get open models hosted in a compliant environment.
OpenAI is teasing Codex hardware with Work Louder. Maybe it is a niche gadget, maybe the start of agent buttons instead of app icons.
🏛️ Claude Fable is back
Anthropic is turning Fable 5 back on after export controls were lifted. The whole incident shows that model access can always be temporary.
The Fable 5 dispute is not gone, according to this Guardian readback; it has just been repackaged. Security review sounds better than a blanket stop, but it is still political access control.
Fable 5 was reportedly significantly faster than Opus 4.8 in a coding test.
Claude Sonnet 5 is available on AWS and aims for Opus-like capability at Sonnet pricing. For many teams, this may become the new cheaper default.
Anthropic is launching Claude Science as a workbench for scientists.
Claude Remote Routines are being described as a work ecosystem: Chat, Code, Cowork, Design, and Routines. The model is becoming a workspace.
A Mythos 6 leak allegedly promises autonomous vulnerability discovery.
🖥️ Google, Microsoft, and AWS are building the agent workbench
Google summarizes its AI updates from June, from Gemini to Android. The common thread: AI is becoming more of an operating layer.
NotebookLM is getting vertical 60-second clips, useful for snackable social media content.
Google's new Home Speaker shows in an early review: the hardware is good, but Gemini as a household assistant still has room to grow.
AWS describes resilience patterns for Bedrock and pairs Nova 2 Lite with Claude for cheaper document processing. Token-Maxxing for enterprise.
🏗️ Infrastructure, money, and power are getting tougher
NVIDIA is expanding the AI factory model with Capital Partners. Jensen is no longer just selling chips, but financing, cloud logic, and revenue participation.
Meta apparently wants to turn excess compute into a cloud product and rent out GPU overcapacity.
The UN warns that AI can worsen global inequality.
Local AI is becoming more attractive as protection against expensive cloud dependency.
Zurich is emerging as a quiet R&D hotspot for the global AI elite. Research follows talent, tax logic, calm, and proximity to major companies.
San Francisco's AI boom makes even $180,000 salaries look small.
Bhavin Turakhia is putting $30 million into Neo, an AI-native Office alternative.
AI stocks are losing ground, according to the Guardian, but the market is still far from a major crash.
🛡️ Society, tools, and the real workplace
Erin Brockovich is taking aim at AI data centers. AI infrastructure is local: water, electricity, land, resistance.
Tripadvisor's AI reportedly softened serious hotel complaints. That is a useful reminder: AI summaries need scrutiny.
A CEO reportedly threatened termination over unreviewed ChatGPT emails. Translation: AI slop is not productivity.
Ford is bringing experienced engineers back because AI cameras missed quality risks.
The Commonwealth Short Story Prize honored a controversial story despite AI-use claims, while e-book lending app Libby is getting an AI filter.
Tidal is cutting royalties for fully AI-generated music.
People who frequently use AI for health topics are more likely to believe vaccine myths, according to a KFF survey.
A US bill aims to ban the sale of health and location data from AI chats.
Rogbid is selling the VisionPro as a $120 AI glasses product with camera and live translation.
Formula 1 teams are wrestling with how to integrate AI sensibly without hollowing out the sport. Better simulation, yes. Invisible autopilot, no.
SpaceX reportedly showed an AI handset prototype; Musk denies it. Everyone is searching for the next interface after the smartphone.
Our thread of the week: AI is growing up, but it is not automatically trustworthy. Whoever manages access, context, cost, and quality well wins. Whoever only chases the strongest model gets stuck again at the next limit.
Done. But no reason to be sad. The AInauts will be back soon with fresh material for you.
See you Monday with a new round of news, workflows, and insights.
Your AInauts
Fabian & Reto
