Pollinger’s Productivity : July 2026

A flood of new models June was relentless. Anthropic shipped Claude Fable 5 – only for it to be ‘banned’ – by the US Government, Moonshot released Kimi K2.7 Code, Zhipu put out GLM-5.2, and right at the month’s end OpenAI unveiled its new flagship GPT-5.6 family – also ‘banned’ by the US Government. To put the pace in context, one industry tracker counted 255 model releases in the first quarter of 2026 alone – roughly three significant models every single day. Insights: You do not need to chase every release; that way madness lies. The smart move is to pick one main tool (e.g. ChatGPT or Copilot), get genuinely good at it, and only switch when there’s a clear reason. If you choose Copilot Premium you get access to GPT models as well as Claude plus more coming soon.

The most powerful AI is now being locked down A striking shift this month: the very best models are no longer simply “released”. OpenAI restricted its new GPT-5.6 Sol to approved users at the request of the US administration during a cybersecurity review, while Anthropic received US approval to bring its top-tier Claude Mythos model back for organisations running critical infrastructure – reportedly over 100 companies and government agencies. Insights: This is a turning point. We’re moving from “newest model wins” to a world where the most capable AI is gated for safety and security reasons. For day-to-day business use it changes little, but it’s a useful reminder that bigger isn’t always better, and that the tools available to you are more than powerful enough for almost any real task.

Copilot Cowork is no longer free When Copilot Cowork went generally available on 16 June, it brought a new pricing model with it, that starts now. You still need a Microsoft 365 Copilot licence to access it, but the actual work Cowork does is now billed separately, on top of your subscription, using a new unit called Copilot Credits. One credit costs about 1p under pay-as-you-go. The number of credits a task uses depends on four things: which AI model it runs, how much data it retrieves, how many tools it calls, and how long it takes to complete. In practice, a rule of thumb is: a simple task – like drafting a status update from your calendar and emails – costs roughly £1–£3. A medium task pulling from multiple sources into a structured document lands around £4 to £7. A heavy task – say, analysing six months of data and producing a leadership summary – runs £7 and up. Credits are shared across Cowork, Copilot Studio, Power Platform, and Dynamics 365 from a single pool. Admins can set tenant-level spending limits and per-user caps from the Microsoft 365 Admin Centre. For organisations that used Cowork during the Frontier preview, the grace period ends today (1 July). From this point, if usage-based billing isn’t configured in your admin centre, Cowork switches off entirely – there’s no warning period. Insights: This is worth paying close attention to – not just as an admin task, but as a signal of where AI pricing is heading. For years, software has been sold on flat per-seat subscriptions. What Microsoft is introducing here is something different: you pay for what the AI does, not just for access to it. A complex, multi-step task costs more than a simple one – much like hiring a contractor by the hour rather than paying a fixed monthly retainer. It’s very likely this is the direction all AI tools are heading. For businesses, that has real implications – both for budgeting (costs will vary month to month based on usage) and for the value question: if you’re paying per task, you’ll want to be confident the output is worth it. The upside is that businesses which use AI heavily and well will see the cost reflect the value they’re getting. Those who don’t use it much won’t be paying for capacity they never touch.

ChatGPT can now run scheduled tasks

OpenAI rolled out Scheduled Tasks letting ChatGPT send reminders, run recurring prompts and quietly monitor things for you, all managed from a dedicated dashboard. Think “every Monday at 8am, summarise this week’s industry news” or “remind me to review the newsletter draft on the 28th”.

Insights: This is the kind of unglamorous feature that genuinely saves time. Start with one small recurring task like a morning briefing, a weekly summary, or a monthly reminder and build from there. It’s interesting that this is being pitched as a new feature. It existed before, but it didn’t work reliably.

Meet Microsoft Scout – the AI agent that works while you’re not watching

Microsoft Scout is Microsoft’s first “Autopilot” agent – a new category of AI that goes beyond answering questions in a chat window. Scout is an always-on expert assistant that works in the background across Outlook, Teams, your calendar and your files. Unlike standard Copilot, Scout doesn’t wait to be asked – it proactively handles meeting prep, flags scheduling conflicts and takes care of routine tasks on your behalf. It’s currently in preview via Microsoft’s Frontier programme.

Insights: The concept of an AI that keeps working in the background – even when you close your laptop – is exactly the kind of tool that compounds time savings over weeks and months. This is a genuine shift in how AI works: from a tool you go to, to an agent that comes to you. When Scout reaches general availability, the question won’t be whether to use it – it’ll be which tasks to hand over first.

💭Dream, dream, dream…ChatGPT just got better at remembering things On 4 June, OpenAI rolled out Dreaming V3 – a completely rebuilt memory system for ChatGPT. The old system stored memories as fixed snapshots that could go stale or simply be wrong. Dreaming V3 works differently: it continuously synthesises and updates what it knows about you from your conversations, keeping your preferences, context and work style current and accurate. The rollout started with Plus and Pro users in the US, and it’s now available in the UK – I had the update today. Memory now includes a helpful Memory summary that you can now update and replaces Saved memories. Insights: This is a more significant update than it sounds. If you’ve ever noticed ChatGPT forgetting your preferences or repeating itself on things you’ve already explained, this is the fix. The practical upside is that your prompts can get shorter over time – you shouldn’t need to keep re-explaining details about you and your business. If you’ve been on my training you’ll know you can pro-actively tailor ChatGPT to your role and set preferred response preferences in Custom Instruction. If you haven’t already set up yours up in ChatGPT (and other tools), hit reply and ask me for my custom instructions Template.