LifeVal: OpenAI puts its models to the test of everyday life, but writes and grades the exam in-house

On November 2, 2026, OpenAI unveiled LifeVal, a benchmark that measures how well a model can get the chores of private life done. The announcement came on the company's official page and in a post from its account on X.

A note on method that applies to the rest of this article: unless stated otherwise, the benchmark's design, the scores and the examples are taken from AI Weekly and CellCog, both of which cite the official LifeVal page. Details reported by only one of the two outlets are flagged.

The scope covers seven areas: caring for family members, health, education, running a household, personal finances, shopping and dealing with public services. According to CellCog, good advice is not enough: the model has to act, meaning browse the web, use the computer and go through connectors. “For AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them,” the company wrote in its post on X (wording ours). LifeVal extends to life outside the office the approach of GDPval, the benchmark OpenAI used in 2025 to measure its models on professional tasks. It arrives as the big labs push so-called agentic assistants, and it shifts the yardstick from the quality of the answer to the outcome of the action.

According to both outlets, OpenAI has used LifeVal internally for about a year to track and improve its models on these tasks. The task list draws on three sources: the American Time Use Survey from the US Bureau of Labor Statistics, the Multinational Time Use Study, and a new survey of 3,995 respondents about where they would like help from an assistant, which the company describes as demographically representative. According to AI Weekly alone, about a hundred participants also logged specific tasks they would have liked to delegate. “To be a truly useful personal assistant, ChatGPT should be able to take actions that get things done,” OpenAI writes on the LifeVal page (wording ours). The published average scores climb from 49.3% for GPT-5.2 to 73.0% for GPT-5.4, then to 77.4% for GPT-5.5, 80.0% for GPT-5.6 and 80.6% for GPT-6 Astra. Among the examples OpenAI gives are enrolling a child in the right preschool after a move (93.1%), coordinating volunteers ahead of a flood (98.0%), planning meals for a parent newly diagnosed with diabetes (90.4%) and watching a game remotely with grandma (96.3%). AI Weekly attributes these scores to GPT-6 Astra's best recorded attempts. CellCog also reports tasks where GPT-6 Astra's best recorded attempt reached 100%, such as booking a car inspection, ordering halal groceries within a budget and unsubscribing from mailing lists. Bear in mind that these are best attempts, not averages.

Plenty of grey areas remain. The task set is not public, there is no leaderboard, and no model from Anthropic, Google, xAI or Meta appears in the results. According to AI Weekly, the evaluation is run and graded by OpenAI with its own systems. In the sources consulted, the company does not say how many tasks there are, what criterion lies behind the reward, how many attempts the average is based on, or whether the tests run on real websites or on test copies. Nor is it clear whether the 3,995 respondents were ChatGPT users, since OpenAI's post speaks of feedback gathered from thousands of ChatGPT users. As for the numbers, the step from GPT-5.6 to GPT-6 Astra is worth six tenths of a point, and the curve seems to be levelling off around 80%. OpenAI does not explain whether the limit lies in the models or in the tasks. LifeVal asks the right question: not just whether the assistant answered well, but whether it actually got the job done. Yet an exam written, taken and graded by the same hand tells you more about the direction than about the finish line. If the tasks are one day made public, or a rival model is put through the same test, that will be the number to read carefully.

— Olya