OpenAI Lets AI Handle Life's Errands, and GPT-6 Astra Scores Several 100%s

On November 2, OpenAI released LifeVal, a benchmark for testing whether AI can handle everyday affairs. A benchmark can be understood as a standardized set of exam questions. LifeVal's questions are very concrete: find a kindergarten, book a car inspection, buy halal groceries, unsubscribe from emails, and make a meal plan for parents newly diagnosed with diabetes.

What It Tests

LifeVal's tasks come from time-use surveys of daily life. OpenAI first used the American Time Use Survey and the Multinational Time Use Study as a foundation, then added a representative survey of 3,995 people to identify the areas where ordinary people most want AI to help. The tasks cover caregiving, health, education, household administration, personal finances, shopping, and government services.

It sets abstract reasoning and coding aside, focusing instead on whether models can get things done in browsers, through computer operation, and via connectors. OpenAI says it has used LifeVal over the past year to track and improve models, with training directions including browser use, computer use, and connectors. The goal is clear: move ChatGPT from a chat tool toward an assistant that can take action itself.

Scorecard

GPT-6 Astra scored 100% on best recorded runs for some tasks:

• Find and book a sensory class suitable for toddlers, 100%
• Check maintenance and book a car inspection before a trip, 100%
• Buy halal groceries within budget and write the order item by item into a spreadsheet, 100%
• Use a photo and a diamond report to find a replacement for a missing engagement ring, 100%
• Find and order a pocket device to replace a phone, e-reader, and handheld console, 100%
• Unsubscribe from mailing lists, 100%

Other tasks were not low either: choosing a kindergarten after moving to a new home 93.1%, managing volunteers before a flood 98.0%, finding a house in Luzerne County 100%, watching a game remotely with Grandma 96.3%, and making a meal plan for parents newly diagnosed with diabetes 90.4%.

How the Perfect Scores Came About

These 100% scores come from best recorded runs. The average across all runs was not provided. DailyAI made the point bluntly in a reply to Mark Kretschmann: 100% comes from selected runs, and he would rather see the average across all runs; OpenAI writes the questions and grades them itself, and that setup is very comfortable.

Kyrelle Buckley asked whether passing childcare means confirming a booking or merely providing a shortlist. It's not an unreasonable question. Between finding a kindergarten and actually securing a place, there are phone calls, availability, deposits, and parents' time. pitching downrange noted that this kind of eval is highly harness-sensitive. How the environment is set up, how browser state is reproduced, and what counts as completion all affect the score.

RAY's reply was more practical: childcare and government services mean the model has to know your entire life, and he wants this kind of capability to run on his own machine—otherwise he doesn't want it. Brad Watton offered a test: if it can handle HM Revenue & Customs (HMRC) over the phone, only then does it count as AGI.

There are positive examples too. Andrew said SOL helped him get a library card last month and even found free audiobooks from his Amazon wishlist at the library. He guesses sign-ups for government programs will rise. Valenciana's complaint also holds up: the benchmark is good, but don't describe it as saving humanity—this isn't life-saving.

One Take

LifeVal's value lies in pulling the AGI narrative back to everyday chores. Finding a kindergarten, buying groceries, unsubscribing from emails, booking a car inspection—these tasks have no benchmark glamour, but they are the work ordinary users would actually be willing to hand over to an agent. OpenAI's page is here: LifeVal.

For developers, the signal from LifeVal is clearer: evals will shift from knowledge Q&A to multi-step, tool-using processes with verifiable results. The real difficulty is how to build the harness, how to reproduce the environment, how to define a pass, and how to calculate cost. For ordinary users, don't be dazzled by the 100%s just yet. Look at the average scores, the failure cases, third-party reproductions, and the privacy boundaries.