On OpenAI's new LifeVal test, GPT-6 Astra scores 80.6% on everyday-life tasks, up from 49.3% for GPT-5.2. OpenAI published the results on Nov. 2. It also built the test and ran it, and OpenAI's post describes no outside verification of the numbers.
What LifeVal measures
LifeVal is OpenAI's attempt to measure how well AI handles ordinary life, not exams or code. OpenAI built it from the American Time Use Survey, the Multinational Time Use Study and a new survey of 3,995 respondents. OpenAI describes that survey as demographically representative. The respondents are a representative sample of the public, not a sample drawn from ChatGPT users, so the tasks reflect what a cross-section of people do, not what people already ask ChatGPT.
The tasks cover caregiving, health, education, household administration, personal finances, shopping and government services. Example tasks include enrolling a child in the right kindergarten, organizing flood volunteers, finding a new home, planning how to watch a game with grandma and preparing meals for diabetes.
The scores
OpenAI reports an "average reward (unweighted)" for each model, a score where higher is better; OpenAI's post as we read it gives no plain-language breakdown of how a task is graded. The figures are OpenAI's.
| Model | Release date | Average reward |
|---|---|---|
| GPT-5.2 | Dec. 11, 2025 | 49.3% |
| GPT-5.4 | Mar. 5, 2026 | 73.0% |
| GPT-5.5 | Apr. 23, 2026 | 77.4% |
| GPT-5.6 | Jul. 9, 2026 | 80.0% |
| GPT-6 Astra | Sep. 3, 2026 | 80.6% |
Most of the gain came early. The jump from GPT-5.2 to GPT-5.4 is 23.7 points. The last two releases add 2.6 points and then 0.6 points, so on this test the newest model is only slightly ahead of GPT-5.6.
On individual tasks, OpenAI shows how "our current models perform" on five example tasks, without naming the model: 93.1% for enrolling in the right kindergarten, 98.0% for the flood volunteers task, 100% for finding a new home, 96.3% for watching a game with grandma and 90.4% for diabetes meal prep. OpenAI also says the best recorded GPT-6 Astra runs hit 100% on some tasks, including booking a sensory class, a car-service inspection, halal groceries and unsubscribing from email lists. OpenAI labels them best recorded scores and does not say how many attempts each took, so they are the high end, not an average.
What this means for you
The pitch is that this is a test closer to real life than most. The tasks come from how people actually spend their time, and the survey behind them was not limited to people who already use a chatbot. That is a sensible design. OpenAI also says it used the evaluation over the past year to track and drive improvement in its models on everyday tasks, so the models were tuned toward this kind of task.
The limits are plain, though. OpenAI built LifeVal, picked the tasks and reports the scores, and the post describes no independent verification. A score of 80.6% is also an average reward, not a share of tasks finished correctly in the real world. We have not tested any of these tasks ourselves.
For more on the model, see our coverage of GPT-6 Astra, Sol and Luna on Google Cloud. OpenAI's other self-run benchmark covered old programming languages, and GPT-6.1 Terra arrived Oct. 26 at a lower price. OpenAI also posted about LifeVal on X.
