The Assistant Is Becoming a Distribution Channel

At DevDay on 29 September, OpenAI announced around 25 things. The three that matter commercially are Dots, which are always-on agents with their own cloud computer and browser that connect to more than 4,000 apps plus Slack and Teams, a marketplace, and Sign in with ChatGPT. Behind all of it, roughly 1.2 billion weekly active users.

Separately, and worth noting for different reasons: GPT-6.1 Sol arrived at close to flagship performance for about a fifth of the price, and OpenAI declined to ship GPT-6.1 Astra at all after deciding it did not meet its safety standards.

What these three features add up to

A place where software gets found, installed, and signed into. That is a platform, and it has the shape of every platform moment that came before it: a large captive audience, a mechanism for third parties to appear in front of them, and an identity layer that makes the whole thing sticky.

The pattern from previous platform moments is consistent enough to plan against. Early participants get distribution they could not otherwise buy. The terms are generous until the platform has enough supply, and then they are not. The categories the platform eventually wants for itself get absorbed, usually the ones with the most volume and the least specificity.

The part that applies to a narrow product

“The platform is a way to be found, not a place to build your moat. Be present with something they would never build themselves.”

Being present in that surface is a distribution opportunity of the kind that does not come often, and it is most valuable now, early, when the terms are good and the supply is thin.

But the thing to be present with matters enormously. A product whose value is general, convenient, and broadly applicable is in the zone a platform eventually occupies itself. A product that encodes the rules of one industry, holds the record for one specific workflow, and is correct in ways that require domain knowledge to even evaluate, is not a category OpenAI will build. There is no version of that which is worth their engineering time, and that is the protection.

The platform is a way to be found, not a place to build your moat.

This is the same conclusion as the Salesforce and Anthropic arrangement pointed at from the other direction. Reasoning is the commodity and the governed specifics are not, so the safe place to stand is the specifics, and the platform is a way to be found rather than a place to build your moat.

The cost curve underneath

Near-flagship capability at a fifth of the price is the kind of move that has now happened often enough to be treated as a trend rather than an event.

The planning implication is straightforward and widely ignored. If your unit economics only work at current model prices, they will work considerably better next year without you doing anything, which is a pleasant position. If your product's advantage is that it uses a better model than competitors can afford, that advantage has a short and predictable life.

The deliberate non-release

The decision not to ship GPT-6.1 Astra on safety grounds is the detail I would pay attention to if I sold into regulated industries.

It establishes that model providers will withhold capability for reasons of their own, on their own timetable. Any product whose roadmap assumes a particular capability arriving on schedule now has a dependency it cannot manage, and the right response is to design so that a delayed or withheld model is an inconvenience rather than a crisis.

See reasoning is the commodity for where to stand, when the buyer is an agent for the transactional side, and what to build when everyone has the same tools for the differentiation question. How the system works covers the positioning.