Top Digital Skills

Ahead of Schedule #09 · September 2026

The month the price went up

OpenAI's new top model costs two and a half times the one it replaced. Why a price going up is the real AI story of September.

Day 1382. 715 days left.

For three years the only direction the number moved was down.

Intelligence got cheaper every quarter. The same answer that cost twenty dollars in 2022 cost seven cents twenty-three months later. Every strategy deck built since then has one unstated assumption buried in it: wait, and it gets cheaper.

On September 3rd, OpenAI shipped GPT-6 Astra at ten dollars per million tokens in, fifty out — about two and a half times what the model it replaced was charging.

The price went up.

That is the whole story this month, and almost nobody is telling it properly, because it looks like a boring pricing note. It isn't. It's the moment the thing you're buying stopped being text and started being finished work — and finished work has never been priced like text.

What moved

1. The frontier got more expensive, on purpose

What happened. GPT-6 Astra landed on September 3rd. OpenAI called it the most intelligent and aligned model it has built, near-perfect on the reasoning benchmarks, state of the art on computer use, software engineering, professional work and science. It also costs 2.5× the model before it, and the first people through the door were companies in an application-only cybersecurity programme — not the general public.

What it actually means. Labs don't raise prices into a competitive market unless they know what the buyer is comparing against. Ten dollars a million tokens is expensive next to last year's model. It is absurdly cheap next to the salary of the person who used to do that task. The pricing has quietly switched reference classes — from software to labour.

What it means if you employ two to fifty people. Every "we'll wait until it's cheaper" plan just expired. The cheap tier is still there and still improving — but the frontier is now being sold to people who have a specific job for it, at a price that only makes sense if you know what that job is worth. Not knowing what job you'd hand it is now the expensive position.

2. The agents stopped needing you in the room

What happened. In February and March, Anthropic, Google and OpenAI each shipped a major model inside a single month, and all three aimed at the same target: long-running, tool-using work that finishes without a human tapping it forward.

What it actually means. The measurable version of this is the one number in AI worth tracking. METR tests frontier models on real tasks and asks: how long a job can this thing finish on its own? In 2020 the answer was seconds. In their January 2026 update, Claude Opus 4.5 completed tasks that take a competent human 320 minutes — five and a half hours of someone's day, done unattended.

What it means if you employ two to fifty people. Five hours is not a prompt. Five hours is a Tuesday morning. The unit of delegation changed from "write me a paragraph" to "here is a job, come back when it's done" — and nothing in how most small businesses assign work has caught up to that. You don't need a new tool. You need a list of the jobs.

3. Agents are arriving inside software you already pay for

What happened. Gartner's forecast: 40% of enterprise applications will have task-specific AI agents built in by the end of 2026, up from less than 5% in 2025. That is an eightfold jump in about a year.

What it actually means. The software you already pay for is growing agents whether you asked for them or not. Most of the adoption happening right now is not a decision anyone made. It's a software update.

What it means if you employ two to fifty people. You may already be running agents. If you are, nobody scoped them, nobody measured them, and nobody decided what they're allowed to touch. The question stopped being should we adopt this about a year ago. It's what is already running here, and is it any good.

The receipt

Post-2023, the frontier's autonomous task horizon is doubling every 131 days. METR's earlier measurement said 165. They re-measured with a bigger task set and the curve had got 20% steeper — not slower. — METR, Time Horizon 1.1, 29 Jan 2026

Read that twice. The people whose entire job is measuring this carefully, with the strongest professional incentive in the field not to overstate it, revised their own estimate toward sooner.

That is what "ahead of schedule" means. Not a vibe. A correction, in the direction of faster, published by the most conservative measurer in the room.

The gap

Here is what most businesses did in the same nine months.

They bought seats. Somebody expensed a Pro subscription, somebody else asked whether it was allowed, and a policy document got written that nobody has opened since. The tool is in the building. The work is unchanged.

That's not a criticism, it's arithmetic. The frontier doubles what it can finish every four months. An organisation's habits change on a scale of years. The gap between those two rates is the entire opportunity, and it widens every single week whether anyone is watching or not.

The businesses pulling ahead this year are not the ones with the best model. Everyone has the same models — that's the point of a frontier. They're the ones who did the unglamorous thing: wrote down what actually happens in their week, and worked out which parts of it are a five-hour unattended job.

That list is the asset. The model is a commodity.

One thing to do this month

Find out what's already automatable in your operation. Not theoretically — in your actual week.

Run the AI Operations Scan — four minutes, free, no call required. It comes back with the specific parts of your operation that are ready now, the ones that need configuration first, and the ones you should leave alone.

Most people are surprised by which is which. The stuff you assume needs AI usually needs a better process. The stuff you never thought about is usually the five-hour job.

Ahead of Schedule is a monthly read on what AI tools can actually do now, and what it changes for a business with two to fifty people in it.

Sources

← All issues