openai’s new middle model costs half as much and makes things up far less

openai released gpt-6.1 terra last week, a mid-range model priced at $1 per million tokens going in and $5 coming out. its api page pitches it at visual reasoning, using a computer and writing that reads naturally, with room for about a million tokens of context. the old gpt-5.6 terra cost $2 and $12. an openai engineer, ted sanders, wrote on hacker news that terra is probably better than its bigger siblings at explaining technical things, about level on vision and computer use but much quicker, and a bit weaker at most other work.

lab claims are lab claims, so here is an outside test. artificial analysis, which benchmarks models independently, found terra scores six points higher than the old terra on its intelligence index while costing about $0.40 per task instead of $1.40. it sits four points behind sol. the biggest jump was in its hallucination test, where terra’s rate fell from 88 percent to 38 percent. on its coding agent index terra scored 48, against 54 for sol and 31 for luna.