OpenAI's third phase puts 'easy enough to use' beside frontier capability

The essay's one operational sentence

Most of 'Built to benefit everyone' is written at the altitude of electrification and lifespan statistics. But buried in the section announcing OpenAI's third phase is a sentence that reads like a product brief rather than a manifesto.

Frontier capability is only part of the job. The bigger task is turning that capability into tools people can actually use to thrive.Montana Labs

That distinction — capability versus usable tools — is the closest the document comes to describing work someone has to ship. The five adjectives attached to the third phase (abundant, affordable, safe, useful, and easy enough) are not model-training targets. Three of them are properties of what a person encounters on a screen: whether the thing is affordable to run, useful for their actual task, and easy enough that they don't give up. That is the frontend.

'Give everyone a personal AGI' is an access problem, not a model problem

The third of OpenAI's three stated goals is to 'give everyone on Earth a personal AGI, empowering them to benefit from one of humanity's most transformative technologies in whatever way they choose.' The examples the authors reach for are deliberately mundane: navigating a medical bill, learning a skill, starting a small business, caring for an aging parent, understanding a legal or financial decision.

None of those are gated by whether a frontier model exists. They are gated by whether a person who is anxious about a medical bill can get to an answer without knowing prompt technique, without hitting a paywall at the wrong moment, and without being confused about what the system just did on their behalf. The essay's own framing — 'whatever way they choose,' 'where and how they need it' — is a description of surface area and access, the part of the stack that lives closest to the user.

The human-role argument raises the bar on interfaces

The authors are explicit that they do not want full automation: 'AI should help people pursue their goals, not become untethered from them.' They describe a growing human role in 'setting direction, making tradeoffs, applying judgment.' They call the automated AI researcher 'steerable, accountable, and connected to people,' and cite an internal belief that a significant fraction of their research may be AI-assisted by March 2028.

Steering, accountability, and connection to people are not achieved in a training run. A system that keeps a human in the loop only does so if the interface actually surfaces what the model is doing, what it decided, and where the person can intervene. The gap between 'the model can do this' and 'the person can direct and correct it' is exactly the gap that frontend and product work closes. The essay asks for judgment and oversight without acknowledging that those depend on what the screen shows.

What OpenAI's framing implies for teams building on top of it

The document's structure — phase one was research, phase two was becoming a product company, phase three is broad usable distribution — is a signal about where OpenAI thinks the remaining work is. It is not primarily in raw capability, which it treats as increasingly assumed; it is in affordability, ease, and the interfaces through which people exercise judgment over capable systems.

For anyone building applications on these models, that reframes the value. If OpenAI itself says the bigger task is turning capability into tools people can use, then the interface layer — how choices are presented, how the system stays legible and correctable, how affordable it is to keep running for someone's ordinary problem — is not decoration on top of the model. In the terms the essay sets out, it is the part that determines whether 'benefits everyone' is true or just aspirational.