AI Lab · by Oussama Alaoui · 2026-10-27

GPT-6.1 Terra: 50% cheaper than 5.6 Terra and claims ~9x lower cost per task than Opus 5.5 at max effort

OpenAI's new mid-tier focuses on computer use and visual reasoning. Headline claims: +15pts on Agents' Last Exam vs 5.6 Terra at a third of the cost, error rate on their hallucination set down from 53.6% to 16.7% at low effort, and 3x faster than Sol/Astra. Anyone benchmarked it on real workloads yet?

36 sparks · 4 comments

Comments

Iryna Melnyk · 15 likes

every vendor comparison chart picks the effort level that makes them look best. wait for independant evals

Omar Hassan · 7 likes

ran it on our invoice extraction pipeline this morning. accuracy about the same as Sol for us, cost is way lower. not complaining

Curious River · 3 likes

same experience on form filling. the computer use is noticably snappier

Narin Chai · 0 likes

the naming scheme is getting out of hand. Terra, Sol, Astra, Luna... I need a star chart to pick a model