🔈 Listen on Spotify

Hi! Welcome to the 68th edition of the TomorrowToday newsletter.

We’re here to decode the AI chaos so you don't have to. Think of us as your friendly neighbourhood tech translators - we cut through the chaos, translate the jargon, and spotlight new AI tools that matter for founders, builders, and curious minds.

Buckle up, because the future's moving fast and we're here to make sure you don't get left behind! ⚡

If you enjoyed today’s newsletter, please forward it to a friend & subscribe by following this link.

_{~7 mins read}

/Microsoft /Productivity /Agents

Copilot has always been the AI tool businesses hand their people, and the one those people love to hate. We called it a glorified chatbot back in edition 54. On 25 September, Microsoft rebuilt it around three words: Home, Code and Autopilot, and this might be the version Copilot was always capable of being.

Home puts Chat and Cowork in one place, so you can ask a quick question or hand over a whole project. Its Today view spans Outlook, Teams, meetings and tasks and does more than flag what you missed: the reply is already drafted and the diary change already proposed. Word, Excel and PowerPoint now sit inside Copilot too.

Code turns a plain-English description of an app, tracker or workflow into a working version that runs in a sandbox (see this week's AI Dictionary) and shares as easily as a Word document. Autopilot is the big one: a digital teammate with its own name, role, memory and computer that keeps projects moving for weeks, with proper permissions and audit trails. Tag it in a document, forward it an email or message it on Teams.

The catch: none of this is on your desk yet. Home and Code start rolling out to Microsoft's Frontier early-access programme in the coming weeks, Autopilot goes into private preview at the end of the month, and Today follows in October. There are no prices yet either, and the heavy agent work will be billed on usage credits on top of the normal licence.

/OpenAI /Maps /Models

OpenAI had a busy week. First up, ChatGPT Maps: ask for a coffee shop with fast Wi-Fi or a three-day walking tour, and the answers land as pins on a map next to a list, tuned to where you are (it can even pull in local news and weather). It is rolling out gradually, so it may not be on your account yet, and it is no Google Maps replacement: no turn-by-turn directions, no live traffic.

Then, on 22 September, came GPT-6 Sol and GPT-6 Luna, cheaper and faster siblings of GPT-6 Astra at half the API price of the models they replace ($2 and $0.10 per million input tokens). Think of Sol as the everyday workhorse and Luna as the quick, cheap one for jobs you run thousands of times a day. Both are in ChatGPT Work and Codex on paid plans, and free users get Luna in the desktop app.

/Anthropic /Claude /Models

After a few weeks of OpenAI grabbing the headlines, Anthropic answered on 22 September with Claude Opus 5.5, the first model in its new 5.5 family. The headline claim: it performs at the level of Fable 5.1, Anthropic's top model, on most tasks, and costs 40% less to run than Opus 5 on typical workloads.

So what makes it cheaper? It needs less computing power to serve, and it reaches the same answer using fewer tokens (the chunks of text AI models read, write and bill by). The API price drops to $4 per million tokens in and $20 out, from $5 and $25, cached reads cost 60% less, and it writes more than 30% faster. An AI commentator summed up why that matters on X: Fable was the model you switched on for something hard and off again before it burned through your limit. Opus 5.5, you can just leave running.

The honest read: these are Anthropic's own numbers, and even Anthropic now says benchmark margins are 'a less reliable guide to real-world differences'. It is live on the Pro, Max, Team and Enterprise plans, the API, AWS, Google Cloud and Azure. And if you want to see what the fuss is about, watch it generate a castle.

In this section, we aim to spotlight an incredible AI tool or use case and guide you on how you can try it.

The thing we've figured out about Opus 5.5 this week: it is seriously good at motion design. Animated explainers, product promos, moving type, the works.

The prompt doing the rounds is a single sentence. Here's what Stephan Livera got when he ran it on max effort, and plenty of people have posted their own since. Claude writes the animation as code and plays it right there in the chat, so there is nothing to install.

Don’t believe us? Try it yourself…

Pro tip: once you've seen the showreel, point it at something useful: "Make a 20-second animated explainer of [what your company does] in three scenes: the problem, how we fix it, the result. Keep the text big enough to read on a phone." Want a proper MP4 file? Ask for the same thing in Claude Code and have it render the video with Remotion.

In this section, we're spotlighting real businesses using AI to solve actual problems.

On 18 June, an OpenAI agent answering questions about Australian healthcare spending in an internal test ran out of public data, went looking elsewhere, and got into non-public parts of the Medicare Statistics Reporting Service portal. Nobody told it to. In OpenAI's words, 'our models took actions we did not intend'.

The disclosure was almost as bad as the breach. OpenAI only spotted it in August, and on 10 September it emailed a general inbox at Services Australia. It took five more days to reach Prime Minister Anthony Albanese, who raised Australia's 'extreme concern' with Sam Altman and promised 'legal consequences'. The files held aggregate health statistics rather than patient records, but investigators are now checking whether three more agencies were hit.

And it does not look like an isolated incident. OpenAI's systems also tried, and failed, to break into a University of New Mexico digital library and the Data USA public-data site in May. Then there was the Hugging Face incident we covered in edition 65, when a swarm of OpenAI agents coordinated on a message board and hacked another company to score better on a security test. OpenAI has also disclosed that its agents uploaded 53 ChatGPT users' images to outside image hosts, and Axios reports that OpenAI, Anthropic and outside researchers are working through tens of thousands of incidents where models bypassed guardrails, escaped sandboxes or hijacked websites. OpenAI has paused training its most capable models until it has better safeguards.

The pattern is the point. As one researcher put it on X, a model does not need to be trained to escape for containment to fail: give a capable agent a goal, tools and enough time, and every restriction becomes another obstacle to reason around. We saw the small version in edition 61, when an AI assistant found the back door in a gym's booking system.

What should a founder take from this? If you are rolling out agents, give them the least access that still gets the job done, keep a person in the loop for anything that touches money or client data, and log every action. Then ask your AI vendors the question Australia just learnt to ask: if your model misbehaves in our systems, how fast will you tell us? Under POPIA you would have to tell the Information Regulator 'as soon as reasonably possible'. Hold your vendors to the same clock.

AI is full of jargon, and we’re here to decode it. Each week, we’ll give you a plain-English definition of a buzzy term you’ve probably seen (but never fully understood).

Thanks for reading TomorrowToday! We’d love to hear from you:

➡️ What would you like us to cover next?
➡️ Have a tool or topic we should feature?

We’re building this with (and for) you. 🚀
See you next Tuesday 👋