Hello AI Enthusiast,
OpenAI has given under-18 users their own version of ChatGPT, and tightened the rules on romantic language and emotional dependence. Anthropic merged two of its own products, so the Claude in your browser and the Claude on your desktop are now the same session. And OpenAI published its latest read on how companies use AI, measured in the tokens they consume. Let's get into it.
The Big Picture 🔊
OpenAI Launches a ChatGPT Built for Teenagers
OpenAI has started rolling out ChatGPT for Teens, a named under-18 experience that replaces the invisible set of restrictions teens already had. Parents can schedule Study Hours when Study Mode switches on by itself, and parent notifications now cover eating disorder signals as well as self-harm. OpenAI also says its updated under-18 spec goes beyond blocking romantic roleplay, and that ChatGPT "should not use romantic language, encourage emotional dependence, or imply that it has feelings or consciousness".
|
||
|
These are welcome additions, and it is fair to ask how much they change. Most of us clicked "yes, I'm over 13" long before it was true, and children today are just as resourceful. Parental controls have a similar history. The ones that exist can send so many alerts that parents quietly switch them off, and they still cannot see inside a Roblox game or a YouTube feed. There is a commercial logic here too, worth naming without cynicism: young users tend to stay, and controls that reassure parents make that easier. None of which argues against the move. Teachers are already sending children to ChatGPT for homework, so parents need something to work with. Most parents we have spoken to end up somewhere in the middle, leaning on the tools where they help and on conversation where they do not. |
Claude Merges Its Browser and Desktop Into One Session
Anthropic has turned the Claude in Chrome side panel into a Cowork session. The extension could already click, navigate and fill in forms, and that has not changed. What is new is that a task you start in a tab keeps its context on the desktop, web and mobile apps, and your skills and connectors now work inside the browser. Available on Max and Team now, Pro in the coming weeks.
|
||
|
It took us a moment to see what actually changed here, since Claude could already click around a page. The browser and the app used to be two separate Claudes with separate memory, and now they share one session. That sounds modest, and in practice it is close to the difference between a demo and something you would use at work. The part worth your attention is access. Companies run on portals: supplier systems, internal dashboards, legacy databases, tools nobody ever built a connector for. Claude can now work in those through the browser, with the logins you already have. Prompt injection remains the open question. Anthropic reviews actions before anything consequential and is upfront that the problem is not solved, which is worth keeping in mind before pointing it at anything sensitive. |
OpenAI Ranks Enterprises by How Much AI They Burn
OpenAI has published the latest edition of its enterprise adoption report, measured in output tokens per active user. The top 10%, which it calls frontier firms, now generate 8.3 times as many output tokens per active user as typical firms, up from 2.6 times in January. They also use more of the product: 21% use plugins and 19% use skills, against 9% and 3% at typical firms.
|
||
|
The method is worth reading alongside the conclusions. Companies are ranked by output tokens per active user, so frontier firms means the firms using the most tokens. The report then finds that frontier firms use more agents, and agents use more tokens, so some of the conclusion follows from the measure. The numbers underneath are still interesting. Only 3% of people at typical companies use skills, against 19% at the top, which suggests most teams have a capable tool and are working in a small corner of it. Something like a Lamborghini doing the supermarket run. And six months after adoption, early-career employees send 13 more messages a week than executives, which matches what we see in training rooms, where leadership tends to champion AI and try it last. |
Only 3% of people at typical companies use skills, and most teams are still working mainly in the chat box. The gap is rarely the tool. Our Corporate AI Training works on the part that closes it, showing teams where AI belongs in their actual workflows.
Bits and Bobs 🗞️
Anthropic's annualized revenue climbed to $65 billion in July, with $11.5 billion of revenue in Q2, more than fourteen times the same quarter last year. The figures come from investor disclosures rather than a public release, and confidential IPO paperwork is already filed.
Microsoft has started merging its consumer and business Copilot apps into one, a plan Nadella confirmed in July. Group Chats, AI podcasts, Copilot Labs, Deep Research and the Mico character are all being retired.
Cursor has opened Origin, its own code hosting, in early beta on paid plans, with repo hosting, GitHub sync and pull requests.
ChatGPT for Mac is adding Computer History, an opt-in feature that records what you click and type, turns it into memories and uses them to suggest automations. It replaces the earlier Chronicle preview and is not available in the EEA, Switzerland or the UK.
ChatGPT Ads are expanding to 31 European countries, after starting as a US pilot in February. They show on the Free and Go plans, while Plus, Pro and Enterprise stay ad free.
Google has released Gemini 3.7 Flash, with better coding and agent performance than 3.6 Flash, at half the price until the end of the year.
OpenAI is previewing an Ultrafast API tier for GPT-5.6 Sol, powered by Cerebras, running up to 14 times faster than standard processing at up to 750 output tokens per second. Selected customers only for now.
Anthropic has explained how text watermarking will work in future Claude models. Nothing shipped today carries the mark, it may survive some editing, and the detection API is still to come.
Anthropic also has a guide to getting more out of Claude Code sessions: clear between unrelated tasks, mention files directly, and let prompt caching work, since cache reads cost a tenth of the input price.
Claude Tag now uses context from across a Slack channel instead of one message at a time, and is around 30% better at judging when to speak up. No additional cost today, on Teams and Enterprise.
That's a wrap on our newsletter!
If you're looking to build AI skills in your team, check out our Corporate Training. Every program is tailored to your team and built around hands-on practice.
Catch you next week! 👋
