OpenAI's dots will switch models mid-task: Terra joins GPT-6 Astra as a 10x cheaper relay

新智元 · wechat · 2026-11-03

OpenAI's ChatGPT/Codex lead Tibo previewed a key upgrade to dots: the personal agent will split a single task into segments and call different models per step, instead of running one model end-to-end.

Terra joins the roster: alongside flagship GPT-6 Astra, Terra ranks #10 of 28 on Vals.ai (60.17%) but completes computer-use tasks 3x faster than Astra/Sol, with API pricing at 1/10 of Astra's.
The logic: mechanical steps (data extraction, form-filling) go to the fast cheap model; the hard reasoning stays with the flagship — saving real time and tokens across a long, dozens-of-steps task.
Risky history: GPT-5's launch-day real-time router broke and OpenAI had to restore the model picker. Dots' multi-model relay within one task raises handoff-failure risks.
What matters: total task time, credit burn, human interventions, and delivery quality. Dots still requires the $100/month Pro plan; cheaper Terra may bring it to Plus users.

Original post →

OpenAI's dots will switch models mid-task: Terra joins GPT-6 Astra as a 10x cheaper relay

More from coding & agent