Laura: OpenAI just dropped a three-tier model family — and the tier names are permanent, not version numbers.
Laura: Sol, Terra, and Luna each have a fixed role. They improve on their own schedule… but you pick the tier for the job, not the latest number.
Laura: The sharpest change is cache billing. Cache writes now cost 1.25x the normal input rate. That's new. Cache reads still get a 90% discount — but there's a 30-minute minimum cache life to qualify.
Laura: So if you're caching aggressively, you need explicit breakpoints and you need to hold that cache long enough to earn the discount back. It rewards planning.
Laura: Then there's Ultra mode. By default it spins up four agents in parallel. Higher token use, stronger benchmark scores, faster time-to-result. OpenAI also tested 16-agent runs on BrowseComp and SEC-Bench Pro.
Laura: Ultra lives in ChatGPT Work for Pro and Enterprise users, and in Codex for Plus and up. In the API, the multi-agent beta inside the Responses API gets you close.
Laura: So the action today — pull OpenAI's published eval tables, run your token volume through their cost calculator, and check whether the Responses API multi-agent beta fits your stack.
Laura: The tier you pick at the start is going to shape your spend and your ceiling for a while.