OpenAI published LifeVal on November 2. It is an internal evaluation that measures how well AI models complete practical, everyday tasks: finding childcare, dealing with government services, making a budget, ordering groceries, and similar jobs. OpenAI says it has used LifeVal for the past year to track and improve its models.

Coffee summary

  • FACT: LifeVal tasks cover caregiving, health, education, household administration, personal finances, shopping, and government services.
  • FACT: OpenAI picked tasks using the American Time Use Survey, the Multinational Time Use Study, and a new survey of 3,995 people.
  • CLAIM: GPT-6 Astra scores 80.6% average reward, up from 49.3% for GPT-5.2, according to reporting on OpenAI’s chart.
  • CLAIM: Example tasks show scores from 90.4% to 100%, such as 93.1% for choosing a kindergarten after a move.
  • FACT: OpenAI has not published the task count, the grading rubric, or the dataset.

What happened

Most AI benchmarks test exams, code, or expert work. LifeVal aims at the tasks ordinary people actually want help with. OpenAI says the tasks rely on abilities like browser use, computer use, and connectors, which move ChatGPT “beyond a chat tool” toward a personal agent.

The OpenAI page shows worked examples. One asks the model to find a home in Luzerne County that fits a budget and a commute, then add the deed to a cart for review. Another plans meals and reminders for a parent with a new diabetes diagnosis.

Why it matters

It shows where OpenAI is heading. If the company is optimizing models for errands, forms, and family logistics, expect ChatGPT agent features to keep moving in that direction.

OPINION: LifeVal is a useful counterweight to hard expert tests. On the same day, SemiAnalysis’s GDPx showed the best agent scoring 12 out of 100 on expert work. Both can be true: AI is getting good at common chores while still struggling with deep professional tasks.

LifeVal examples, at a glance

Example task Score shown by OpenAI
Find a home that fits budget and commute 100%
Organize volunteers before a flood 98.0%
Set up a remote game-watch with grandma 96.3%
Enroll in the right kindergarten 93.1%
Meal plan for a parent with diabetes 90.4%

Who should care

  • Everyday ChatGPT users wondering what agent features will be good at.
  • Product teams building assistants for consumers.
  • Researchers who track how labs measure progress.

Limitations and open questions

  • Self-graded. OpenAI built the test, chose the tasks, and reports its own models.
  • No rivals. No Claude, Gemini, or Grok scores appear.
  • “Average reward” is undefined. The page does not explain the rubric or how many attempts sit behind each number.
  • Best, not typical. Some 100% results are listed as best recorded scores, not averages.
  • Not public. There is no dataset or code for others to rerun.

What to do next

  1. Treat the numbers as CLAIM until outside groups can test them.
  2. Try a real errand in ChatGPT agent mode and check every step before it submits anything.
  3. Never let an agent pay or sign without your review.

AIImpish Take

LifeVal is more of a progress chart than a scientific benchmark, but it is an honest signal of priorities. OpenAI wants ChatGPT to run your errands. Useful, as long as you stay the one who clicks “confirm”.

Related on AIImpish

*Sources: OpenAI, LifeVal; AI Affairs, OpenAI introduces LifeVal; Neural Brew, OpenAI Unveils LifeVal.*