TWP: The Update Strikes Back

Hi friend,

Two flagship AI models landed in the space of two days this week. Both came with fine print about what your plan includes, and in both cases the fine print is the story.

OpenAI released GPT-6 Astra on Thursday. It is built for work that has a right answer. Driving software on its own, writing code, maths and science, and producing documents, spreadsheets and slides that hold to a template you hand it. OpenAI says it finishes computer-use tasks in roughly half the time the previous model needed, and it gets there using about a third of the words (OpenAI). Give it a messy folder of numbers and a format to follow and it will do better than anything you had last month.

Writing is the part to watch. Independent reviewers put Astra below GPT-5.6 Sol, the model it replaces, on editorial writing, describing the output as flat and easy to spot as machine-made (BigGo). If you have a prompt that currently produces copy you are happy with, swapping the model underneath it is the quickest way to break it. Keep the writing where it is and give Astra the spreadsheet.

Then the rollout came apart. The announcement said Plus, Pro, Business and Enterprise. What Plus subscribers found was Astra sitting inside ChatGPT Work and Codex while the ordinary chat window still served up GPT-5.6 Sol. Sam Altman apologised the next morning for a "messy" rollout (The New Stack), and OpenAI began banking one usage reset per day for everyone still waiting (Implicator). By the weekend some Plus accounts still only had it in Work (BleepingComputer).

Open your model picker before you believe anything written about what you can now do, this newsletter included. If Astra is not there, check the Work section. When it does arrive it runs inside your existing subscription limits, so there is nothing extra to buy.

Anthropic put out Claude Fable 5.1 the day before, aimed at the opposite kind of job. Not quick answers, but long ones: multi-step research, sustained coding, chewing through a stack of documents without losing the thread halfway. It also says no far less often, blocking 60% fewer false positives on cybersecurity questions than the version before it (Anthropic). Anyone in security, finance or health who gave up on Claude because it kept refusing perfectly ordinary requests has a reason to try again.

The cost of that is patience. Fable 5.1 thinks for longer and writes far more to reach an answer, so a single task takes longer and costs more than the same task on Astra (Artificial Analysis). Mythos 5.1 landed alongside it, the same model with lighter safeguards, and you cannot have it: it goes to vetted cybersecurity and life sciences organisations only.

Its fine print costs actual money. On the Pro plan Fable 5.1 shows up in your model picker but sits outside your usage limits. You pay for it with usage credits on top of the subscription, and unlike the July change to Fable 5 there is no one-off credit this time to soften the landing (Anthropic). Max plans and Team premium seats get it included, up to half of the weekly limit. Free accounts do not get it at all.

Worth two minutes if you pay for Claude: look at which model is selected right now, then look at your credit balance. A picker that offers you a model is not the same as a plan that covers it.

Everyone got one thing this week, free accounts included. ChatGPT’s image tools were rebuilt on Tuesday and shipped globally on every plan (OpenAI). There are templates under Images, a Sketch option on mobile where you draw the rough idea and describe the rest, and a full-screen view where you comment on the picture to change it instead of rewriting the whole prompt. You can also share the prompt behind an image so someone else can run their own version. Generation limits have not moved, and templates have not reached Work mode yet.

Try this one on something you actually need:

“Generate a 16:9 header image for a post about [your topic]. Photographic, natural light, no text, no logos, no people. Before you render anything, describe three different visual ideas in one line each and let me pick one.”

Smaller things, quickly. Google Vids will now turn a Doc, PDF or Word file into a narrated video summary, though it needs Business Standard or higher (Google). Custom Gemini instructions now follow you across Gmail, Drive, Chat, Sheets and Slides rather than resetting in each one, which makes them worth writing properly once. Google Pics image generation reached general availability on the same tiers. Sheets charts got the full Google Fonts library, which sounds trivial until you have tried to match a chart to your brand. Microsoft’s Copilot release notes have not moved since 25 August (Microsoft), so if your Copilot feels the same as it did a fortnight ago, that is because it is.

See you next week,

Shantanu

P.S. Forward this to whoever in your office argues that the AI subscription is a waste of money. (They may be right. Check the credit balance first.)