Why OpenAI's cheaper models matter more than its star
OpenAI's new top model set a record. But the bigger deal is that near-top AI now costs about half what it did — and that changes what's worth automating.
- ~50%
- What Terra costs vs. the old top model, GPT-5.5
- 3
- models, priced high to low: Sol, Terra, Luna
- 91.91%
- Sol's record score on the Terminal-Bench 2.1 test
- ~20
- companies let into the limited preview
OpenAI showed off new AI models called GPT-5.6 on June 26. Everyone repeats one number: the top model, Sol, scored 91.91% on a hard coding test called Terminal-Bench 2.1 — the best score any AI has ever gotten. But the real news is the price. The cheaper models do almost the same work for about half the cost.
There are three models. Sol is the star. It can think longer and step by step in a "max" mode, and split a job across several small helpers at once in an "ultra" mode. The middle one, Terra, is just as good as the previous best model, GPT-5.5, but costs about half as much. The smallest, Luna, is the fastest and cheapest, made for simple jobs you run over and over.
This matters if you build software with AI. Most everyday jobs — sorting text, pulling out facts, writing first drafts — don't need the smartest model. Use the cheapest one that does the job well, and save the pricey star for the truly hard cases. Same result, a fraction of the cost. That can be the difference between an AI feature that makes money and one that quietly loses it.
One catch: you may not even get in. The preview is open to only about 20 companies checked by the U.S. government. OpenAI didn't like that rule and said such government access shouldn't "become the long-term default." An earlier version of the rule forced a rival to pause its newest models. So don't lock yourself in. Build so you can swap one model for another, and expect the cheapest good-enough model to keep changing.