The family makes model selection a workload decision, alongside the amount of reasoning requested.
What happened
OpenAI reports stronger coding, knowledge work and computer use, with improved efficiency. It also introduced an ultra setting that coordinates multiple agents on demanding projects.
Why it matters
A product can send different jobs to different tiers, reserving more expensive computation for cases that benefit from it. Claims about performance per dollar depend on the task and evaluation setup.
The bigger picture
Our take. Measure successful completion, elapsed time and review effort together. Lower token consumption is useful, but a cheaper answer that needs extensive repair can still cost more overall.
Source checked November 18, 2026. Company claims are attributed; paragraphs marked “Our take” are editorial interpretation.