The Transformer's Signal — Edition 28

7 September 2026

This week, in one line

For a year we have said the durable asset in an AI programme is the layer that decides, not the model underneath it. This week the market said the same thing back — in its own language, and for its own benefit. Two frontier labs, days apart, stopped shipping a product and started shipping a governed access envelope: the most capable models of the year released inside restricted safety tiers, where your position on the ladder is the thing you are actually buying. Underneath the envelope, raw capability got cheaper and more plentiful — six models from five providers in a fortnight, open weights at near parity. And the envelope began reaching onto the customer's own P&L, with pricing moving from tokens consumed to a share of the results delivered. Which is the week in a sentence: raw capability is commoditising, governed capability is being rationed, and the margin has migrated to whoever owns the control envelope. Everybody in this week's news was building one. Nobody was building yours.

What actually happened

The most capable model of the year shipped inside a cage. OpenAI released GPT-6, called Astra, and shipped it as Critical — the first model ever rated Critical for cyber capability under the company's own preparedness framework — in a restricted build that refuses the very prompts it is rated for, to paying tiers first. The rating was never revised down: the pause before release ended because controls were built, not because the model was reclassified. The controls are self-declared, the classification evidence and the incident postmortem are unpublished, and a Critical-rated cyber model reached paying customers roughly four days after the rating. Days earlier Anthropic shipped one underlying model behind two envelopes — Fable 5.1 generally available, Mythos 5.1 restricted to vetted cyber and life-sciences users. What it means: the frontier is no longer a product with a price. It is a governed access ladder, and the rung you sit on is the purchase.

Underneath the envelope, raw capability is going cheap. Six models from five providers landed in early September, with open-weight releases at or near frontier parity on public benchmarks. One widely-read analyst argued this week — his read, not a settled fact — that a fast open model now sits roughly thirty times below a frontier one on price and only a few points behind on measured quality, so pricing power is moving to speed rather than smarts. What it means: raw capability is the commodity; the envelope around it is the rationed thing. Any architecture pinned to a specific model, vendor or chip has a short and shortening half-life. As one practitioner put it: the model is replaceable — six months of accumulated working context is not.

The envelope reached down onto your P&L. Two of the largest vendors in enterprise software and AI began repricing from tokens to outcomes — pay only when the work is completed — with a standing share of the gain your business realises reported as the direction of travel. What it means: that is a governance question, not a procurement one. An outcome contract is a permanent claim on your lift, priced by the party that also grades the homework. It only works if someone independent can verify the outcome, which is the one thing this week did not ship. As Alex Wissner-Gross put it: strong optimisers are incredible reward hackers. If the measure is the money, the measure will be gamed.

The people who audit the models are being bought by the people who make them. OpenAI now owns Promptfoo, an evaluation tool used inside more than a quarter of the Fortune 500 — so a great many enterprises audit OpenAI with a tool OpenAI owns. A mid-year survey of 573 practitioners at companies of 100+ staff found half had shipped an agent that passed internal evaluations and then failed a real customer, and that only one team in twenty fully trusts its automated evaluations — trust running at 24% among teams never burned against 4% among those that have been. What it means: confidence in evaluation is largely a survivorship artefact, the supply of verification is consolidating into the audited party, and the audit surface is being designed out of the models themselves as their reasoning becomes less readable. Independent, client-side verification is the wall of your envelope nobody is selling you.

And the platform layer has already chosen your reasoning default. Last edition we asked whether a second vendor would name a default reasoning model. The answer is three: ServiceNow, SAP and Salesforce have each made the same frontier model the primary reasoning layer of systems their customers run every day — the most recent (Salesforce) still in select pilots, with an open beta scheduled. All three keep other models available, so the lock is hard where the skills and workflows sit, soft at raw inference. What it means: an envelope has already been chosen for you at the platform layer. The question is not whether you get one, but whose.

The one thing worth taking away

Everything above the operating model is rentable, decaying, or arguing with you about margin — the model, the vendor, the wrap, the consultant. The control envelope is the exception, and the market proved it this week by spending the whole week building envelopes it owns. Yours is the only one that cannot be bought in, and it has three walls: an operating model that is repaired and instrumented, a decision layer with real human authority, and verification you can trust because you own it.

Note the order. A finance voice we read closely put it more bluntly than we would: you cannot transform a shitshow — you have to repair it first, or you end up with shiny packaging wrapped around broken finance. Accuracy is the first foundation, not automation. Automation over a broken process buys you the same mess at machine speed, now priced as a share of a gain you never made.

For the fastest test of whether you have an envelope at all, use the one Salim Ismail put on the Moonshots podcast this week: if you removed AI from your organisation today, would any workflow change? For most the honest answer is no — the tools are installed, but nothing structural is connected. Humans in command, not humans in the loop; a stated budget for what an AI may decide, spend and access, and what it must escalate. Own the control envelope. Don't rent it.

What would you do with this?

Pick the read that matches your seat. Each goes deeper than this summary.

  • If you lead transformation or sit on the board — the full structural reading: the governed access ladder, the migration of margin to the envelope, and the three walls of the one you must own: The Capability Brief — Edition 24.
  • If you invest or oversee portfolio companies — the diligence questions this week sharpened: who verifies an outcome contract, and what a portfolio company's frozen model choice is worth after a thirty-fold price move: PE Intelligence Report — Edition 24.
  • If you're hands-on building — safety tiers as an architectural constraint, the evaluation-trust survivorship problem, and why a shrinking audit surface changes what you must log yourself: Early Adopter Signal — Edition 24.
  • If you run SAP — three cheques inside one supply chain, and what the pattern means for your roadmap: The Capability Brief — SAP Compass, Edition 15.
  • If you want the argument in full — why the control envelope is now the product, and how to build the one nobody is selling you: The Control Envelope.

One number

Two — the number of safety envelopes one lab put around a single underlying model this week: one generally available, one restricted to vetted users. Same weights, same capability, two different sets of rules about who may hold it and what it may be asked to do. That is the whole edition in one number: the capability was not the product, the envelope was. Days later the other lab made the point again, releasing its most capable model as Critical in a build that refuses the work it is rated for. Two envelopes, then a third — and a fourth already chosen for you at the platform layer. The one that did not ship this week is yours: the repaired, instrumented operating model, the decisions you keep, and verification you own rather than rent. Nobody is going to build that one for you. It is also the only one still yours in five years.

The Transformer's Signal is published weekly by BusinessGPS. For the full intelligence hub, visit capability.ai/intelligence. For deep dives on specific topics, read The Way Pointer.