OpenAI grades its own AI on real-life tasks. GPT-6 Astra hits 80%.
OpenAI launched LifeVal to test AI on chores like grocery shopping and budgeting, claiming GPT-6 Astra beats older models by 30 points.
Models · Source: CNN
What happened
OpenAI just dropped a new benchmark called LifeVal. It measures how well AI handles practical, everyday chores instead of just academic tests. Think finding childcare, creating budgets, and coordinating flood volunteers. For the past year, OpenAI has used this internal tool to monitor model results and direct improvements. The goal is to see if a model can help someone make decisions and carry out meaningful actions.
To build the test, OpenAI looked at government time-use surveys from the United States and the United Kingdom. They supplemented that data by polling 3,995 users to find out what people actually wanted an AI assistant to do. Then they had a smaller group of about 100 experts write the grading rubrics. These experts had to prove they understood what a correct answer looked like for complex tasks like financial planning.
The company claims its GPT-6 Astra model, released in September, scored 80 percent on these everyday tasks. That is a massive jump from the 49.3 percent scored by GPT-5.2 in December 2025. On specific workflows, the numbers are even higher. It reportedly hit 100 percent on a Luzerne County home search and 93.1 percent at comparing kindergartens after a household move.
Key facts
- 80% — GPT-6 Astra score on everyday task prompts
- 49.3% — GPT-5.2 score on the same benchmark in December 2025
- 10 million — Downloads for Meta's Muse agent in its first 40 days
- 3,995 — Survey respondents used to build the LifeVal evaluation
Why it matters
The battleground for AI companies is shifting from chatting to acting. OpenAI, Meta, and Anthropic are all racing to build personal agents that execute tasks in the background. Meta already saw 10 million downloads in 40 days for its Muse agent. Anthropic announced Claude Duet in October. OpenAI introduced its Dots agent in September to work on long-running tasks. If models can reliably navigate websites and fill out forms, the value of AI shifts entirely. It moves from generating content to giving people their time back.
But benchmarks are rapidly becoming marketing tools. OpenAI built the LifeVal test, picked the tasks, and graded its own models. Measuring how AI performs in everyday life is highly subjective. The steps required to shop for groceries or pick a kindergarten are broad and vague. Until independent developers see these agents successfully complete multi-step actions in the wild without hallucinating, these high scores are just directional signals. The real test is not a lab score. The real test is whether the AI actually orders the right groceries and handles the payment without failing.
For builders
Build connectors for everyday services
OpenAI explicitly stated its models became more aligned for browser-use and connector tasks over the past year. Builders who create reliable APIs for local services like grocery delivery or apartment hunting will capture the traffic these agents generate. The platforms that integrate easiest will win the automated spending.
Do not trust vendor benchmarks blindly
OpenAI claims a 90.4 percent success rate on diabetes care planning. Founders building health or finance apps must run their own rigorous evaluations. A generalized benchmark from the model creator does not guarantee safety or accuracy in high-stakes vertical use cases. You carry the liability, not OpenAI.
Shift focus to agentic workflows
The product direction for ChatGPT is moving toward an agentic personal assistant. Engineers need to design applications that expect to interact with AI agents rather than human users. This means prioritizing machine-readable interfaces and deterministic execution over flashy user interfaces.
My take
AI can easily pass a coding exam, but it usually fails at buying milk. OpenAI knows this gap exists, which is why they built LifeVal to measure practical assistance. But grading your own homework is easy. I will believe the 80 percent score when the AI actually manages my calendar and buys my groceries without breaking anything.
Original reporting: CNN. This is my rewrite and opinion.