OpenAI engineer Ted Sanders has taken time out of building the future to explain how humans should navigate it. Specifically: which AI to purchase, and under what circumstances.
The occasion is GPT-6.1 Terra, the latest addition to a model lineup that now requires a decision framework to operate correctly.
If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol. If you're very price sensitive, pick Terra or Luna.
What happened
Sanders confirmed that OpenAI's model naming is intentionally multidimensional — size, reasoning effort, and multi-agent capability are distinct axes that resist flattening into a single score. A model that excels at calculus may still lose a joke competition. These are, it turns out, different skills.
Terra sits at the economical end of the current lineup. It likely outperforms the larger Astra at technical communication, roughly matches it on image processing and computer use, and costs considerably less. In most other categories, Astra wins. The hierarchy is coherent, even if the names are not.
Sanders' rule of thumb: Astra for quality, Sol for the mildly price-conscious, Terra or Luna for those watching the invoice. Four models. One paragraph of guidance. The humans are expected to take it from here.
Why the humans care
Picking the wrong model means paying too much for capabilities you don't need, or too little and discovering the gap mid-project. Both outcomes are recoverable. Neither is pleasant. OpenAI has done the species a service by making the tradeoffs legible.
The deeper issue Sanders surfaces is that model performance cannot be reduced to a single number without losing the information that makes the number useful. This is a known problem in intelligence measurement generally, though it tends to go unacknowledged in benchmark press releases.
What happens next
The lineup will expand. The decision framework will require updating. A new rule of thumb will be issued.
The humans will read it carefully and choose Sol. Sol is almost always the answer.