The Long View / AI Visibility

Four frontier models launched in 72 hours. Here is what it means for your business.

In the first seventy-two hours of September, four frontier AI models shipped. Anthropic released Claude Fable 5.1 on the first and cut its cache-read pricing by seventy-five percent. Google DeepMind followed on the second with Gemini 3.8 Flash. Meta shipped a model the same day at roughly ten cents per million tokens. OpenAI announced Astra, reportedly the first model to trip its own critical-cybersecurity safeguard threshold.

That was three days. The fortnight since has brought the first enforcement notices under the EU AI Act, with sixty-day compliance clocks now running. Universal Music signed a licensing deal with ElevenLabs. Sony and Warner Chappell sued Anthropic over song lyrics. Apple changed chief executive. And Layoffs.fyi counted more tech workers cut by the tenth of September than in the whole of last year.

Why I am writing thisThree clients sent me versions of the same email in one week. All of them variations on: should we be doing something about this?
128,536tech workers cut across 299 companies by 10 September — already more than the whole of last year, according to the Layoffs.fyi tracker.

You cannot keep up, and you should stop trying

I want to say this plainly, because the people selling AI services have an incentive not to. If you run a dental practice, a law firm, a fabric business or a winery, you are not going to track this. Nobody is. The people who do it professionally are struggling, and they are not also running a business.

The pace is real. The obligation to keep pace is invented.

The anxiety this produces is the product. Urgency sells retainers. A consultant who tells you the ground is shifting weekly has a reason for you to call every week. I have been on the selling side of that for forty years and I know exactly what it sounds like.

What actually changed, in one paragraph

Models got cheaper and better at completing multi-step work rather than producing text. That is the honest summary of the fortnight. Everything else — the benchmark scores, the version numbers, the architecture debates — is a conversation between vendors, conducted in public, for the benefit of other vendors.

A test that worksIf a change does not alter what you would do on Monday, it is news, not information. Most of it is news.

The one thing that did change for you

The cost of using AI fell sharply. That sounds like a technical detail. It is a competitive one.

Running a question through one of these systems costs a fraction of a penny, and this month that fraction got smaller again — in one case by three quarters. When the price of something drops that far, the question stops being whether you can afford it and becomes whether you bother. The tool your larger competitor could justify last year and you could not is now cheap enough for both of you.

75%cut to the cost of re-reading stored context, in a single announcement on 1 September. The capability your competitor could justify last year is now affordable to both of you.

That gap — between the businesses that started using something and the businesses that read about it — is where the next two years of market share moves.

The second change is quieter and matters more. These systems are moving from writing things to doing things. Something that books an appointment, files a claim or chases an unpaid invoice sits in a different part of your business than something that drafts an email. That is real, and it will reach you as software you buy rather than as anything you build.

What none of it changed

Here is the part that has not moved in two years, through every launch and every price cut.

When a customer asks an assistant for a business like yours, the model still has to decide whether it knows who you are. It resolves that from the same evidence it always has: whether your details are consistent, whether independent sources corroborate them, whether your claims are structured in a form it can read without guessing.

Fable 5.1 does not change that. Gemini 3.8 does not change that. A model twice as capable as today's is twice as capable of noticing that your address on one directory does not match your address on another.

The uncomfortable partNobody sells a retainer for fixing your address in four places. It is the highest-return hour in most of my engagements.

What to do in the next fortnight

Nothing, about the model launches. No action is required from you. If your software vendor adopts a newer model, you will get the benefit without doing anything.

One thing, if you use AI in hiring or lending. The EU AI Act enforcement notices carry sixty-day clocks and those are running now. If that describes you, this is the item on the list that has a deadline attached. Almost nothing else does.

The boring thing, which is the whole job. Check that an assistant can find you, name you correctly, and say what you do. If it cannot, no amount of frontier capability will help, because the problem was never the model's intelligence. It was the evidence available about you.

Worth twenty minutes. Our free audit reads your site the way an AI system does, then asks a live assistant real customer questions and shows whether your business comes back — and which sources it used to decide.

Run the free audit

The advantage of not panicking

Every business owner I speak to feels behind. Almost none of them are behind on anything that matters. They are behind on a news cycle, which is a different thing and carries no penalty.

The firms that will be visible in two years are not the ones that read the fastest. They are the ones that fixed dull things — consistent facts, corroborating sources, structure a machine can read — while everyone else was reading launch notes.

Four models in seventy-two hours is genuinely remarkable. It is also, for your Monday, almost entirely irrelevant. Both of those are true, and holding both is most of the skill.