Thursday at ten the shutter still goes up. Maps still has last summer’s hours. The eight o’clock WhatsApp is still sitting there. On the phone, three names. Fable 5.1 has been in Cursor since 1 September. Astra — OpenAI’s GPT-6 — shipped today. Grok 4.7 is the one people expect in about eight to ten days. The picker fills. The door does not.
Not a named client. A workshop, a small hotel, an optician that already bills and can pay to do this properly. Forget which model is cleverest. Ask what changes behind the counter.
What a shop that can pay actually gets
Almost nothing the feed is selling. The useful bit is dull: cleaner drafts, cheaper long sessions.
Fable 5.1 is the one we already run for work that lasts hours — the same card, the same menu, read again. Cursor’s docs give one number I will repeat: prompt-cache reads are 75% cheaper than on Fable 5, $0.25 against $1.00 per million tokens. Input and output stay put. I will not invent a studio saving on top of that.
That lands in three places. The site — card, page and door saying the same thing. A page that takes the booking, not the chat used as a diary. And a workflow with a person in it: the agent drafts the confirmation in the language it arrived in, someone reads it, someone sends it. Better draft. Cheaper session. Same judgment on the last door.
Astra is today’s name. It shipped — treat it as live — but in stages. Daybreak, the cyber-defence build, went first. Plus, Pro, Business, Enterprise, the API and AWS are “in the coming days.” A shop in Vélez does not have it in the till this morning. OpenAI points it at software engineering, computer use, professional documents and spreadsheets, multistep agent work. None of that opens your shutter. And it is not in Cursor. This morning the list is still Fable 5.1, Grok 4.6, Grok 4.5, Composer, Opus. Not Astra. The name landed outside what we run. We look at it the way we look at any new picker entry: it drafts, a person sends.
Steal this if it fits
We run several models at once, Cursor plus a Grok Bot. No secret. If the pattern fits a shop, take it.
The long stuff — a sector kit, the first three of a wave, Spanish that has to land — goes to Fable 5.1, where that cheaper cache actually shows. When the tone cannot wobble, an Insights note, a client email, Opus 4.8 writes it. Before a person signs, Grok reads the draft and says go, fix or hold. Names, towns, the boring fills: GLM. French waits until English and Spanish are locked; Composer does that, and the implementor chores. Four Grok Bots — Desk, Copy, Reviewer, Scout — each draft their lane. Same rule on all of them. They draft. A person sends. Never on their own.
That is what survives a new name in the picker. Not the model. The habit: something drafts, someone signs.
When 4.7 shows up
It has not shipped. Cursor still lists 4.6 and 4.5. The signal we have is about eight to ten days. No lab score. No date nailed to the wall.
We evaluate it then. We do not reshuffle tonight. The seat is the one Grok already has: the go/fix/hold gate, and the strategy. If 4.7 is sharper there, it gets tried there. Not everywhere at once. Your shop, this morning, still has a door. No feed name opens it.
The close
The picker will keep changing. Don’t build the business on one name. Build the workflow — something that drafts, someone who sends — and let the models sit underneath.
When we do that as a package, it comes in three sizes. Presence (€3,450) gets the card, the site and the door telling one story. Bookings (€5,450) puts the booking on your own domain, with WhatsApp as backup. Operator (€7,950) adds one real workflow — an agent that drafts in the language it arrived in, and a person who reads it and sends it. Care (€149/month) keeps the hours, the slots and the fixes current; it is part of the work, not an add-on you drop to save money. Named once, here, so the shape of the work is clear.
Ask about a Web & eCommerce package. A person replies by email within 24 hours.
Frequently asked questions
Do I have to switch models every time a new one lands?
No. The picker will keep changing names — don't build the business on one. What holds is a workflow where the model drafts and a person reads and sends; that survives even if the best name is different tomorrow.
Are Astra or Grok 4.7 in Cursor yet?
Fable 5.1 is, since 1 September. Astra — OpenAI's GPT-6 — shipped today, but in stages and not in Cursor: this morning the docs still list Fable 5.1, Grok 4.6, Grok 4.5, Composer and Opus, not Astra. It landed outside what we run, and we look at it the way we look at any new picker entry: it drafts, a person sends. Grok 4.7 is expected in about eight to ten days and hasn't shipped. In your shop the difference is still the door, not the model name.
Will the new model answer my customers on its own?
No, and we don't want it to. An agent drafts the reply or confirmation in the language someone wrote to you; a person reads it and sends it. Nothing goes out alone. What the newer models make cheaper is the draft, not the judgment.