OpenAI says its new benchmark puts GPT-6 Astra at 80% on everyday tasks, against 49.3% for GPT-5.2. The result gives a concrete measure of the company’s push to make ChatGPT handle errands and planning, though it is a company-designed evaluation, not a universal report card for AI assistants.
OpenAI Watch analysis
What happened
OpenAI introduced a benchmark for tasks such as shopping, finding a school and planning meals. CNN reports that the company surveyed around 4,000 people about what they wanted an AI assistant to do, then asked a smaller group of about 100 to propose tasks, provide relevant personal context and define what a successful result would look like.
OpenAI says GPT-6 Astra scored 80% on general everyday-task performance. CNN also reports task-specific scores claimed by the company: 93.1% for finding a suitable kindergarten after a move, 98% for managing community volunteers preparing for a flood and 90.4% for making a meal plan for a parent recently diagnosed with diabetes.
Why it matters
The benchmark is aimed at a real gap between an assistant that produces a polished answer and one that can navigate websites, use information from a person’s chat history and reach a result that fits their needs. Those are the capabilities needed if ChatGPT is to move from generating content towards doing useful work on someone’s behalf.
OpenAI’s figures are encouraging, but the tasks and scoring rubrics were developed through its own process with users. Everyday needs can be subjective, and a strong score on selected tests does not establish how reliably the model will handle other people’s circumstances.
Our read
This is more useful than another vague promise that AI will “save time”: OpenAI has given a headline score and described how it assembled the tasks behind it. But the meaningful test is whether those results hold up on tasks chosen and assessed independently. A tidy benchmark is a start, not a substitute for asking who chose the exam questions.
What to watch
- Whether OpenAI publishes the tasks, scoring rubrics and more detail on how the scores were calculated.
- How GPT-6 Astra performs on independently designed everyday-task evaluations.
- Whether users can see and correct the personal context an assistant relies on when taking action.
Discussion spark: Should AI companies be allowed to define and score their own everyday-task benchmarks, or should independent evaluators set the tests?
Sources and evidence
- OpenAI says its AI is getting better at doing things humans do, like shopping and meal planning – CNN (2 November 2026, 17:06 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.