OpenAI unveils an everyday-tasks benchmark and says GPT-6 Astra scores 80%, up from 49.3% for GPT-5.2
CNN reports the internal evaluation was built from about 4,000 surveyed ChatGPT users; OpenAI has not yet published the test set.
Why it matters
OpenAI is extending its evaluations from paid work, measured last year by GDPval, to household errands such as shopping, apartment hunting and caring for relatives. The reported jump from 49.3% for GPT-5.2 to 80% for GPT-6 Astra is OpenAI's own measurement, with rubrics written by about 100 users who proved expertise in each task. It arrives as Meta's Muse, OpenAI's Dots and Anthropic's Claude Duet compete as personal agents. Watch for a public task set and independent replication.
Key facts
- GPT-6 Astra: 80% overall on everyday tasks
- GPT-5.2: 49.3%
- About 4,000 ChatGPT users surveyed
- About 100 users wrote tasks and rubrics
- Kindergarten search task: 93.1%
- Diabetes meal-plan task: 90.4%