OpenAI reportedly releases LifeVal life-task benchmark, disclosing scores across multiple generations of models

AT A GLANCE

Reports say that OpenAI has released LifeVal, a life tasks benchmark, in which GPT-6 Astra scored an average of 80.6%, but the test was designed and judged by OpenAI itself.

Article

OpenAI Releases LifeVal: From Grocery Shopping and Finding a Home to Caring for Family Members, Specifically Testing Whether AI Can Handle Everyday Life

Dongcha Beating AI News, OpenAI has released a new benchmark called LifeVal, specifically designed to test whether AI can help ordinary people handle everyday life tasks. From finding kindergartens, grocery shopping, applying for government benefits, to caring for sick family members, the tests all involve real-life situations that require step-by-step handling.

To determine the test content, OpenAI referenced survey data on daily life and time allocation of residents across various countries, and recruited 3995 people to ask them what tasks they most wanted AI to help with. LifeVal thus covers areas including childcare, healthcare, shopping, financial management, and education.

The results published by OpenAI show that GPT-5.2 achieved an average score of 49.3%, while GPT-6 Astra has improved to 80.6%. However, the progress was mainly concentrated in the early stages: GPT-5.4 had already reached 73%, GPT-5.6 was at 80%, and the latest Astra is only 0.6 percentage points higher than it.

The official also showcased some high-scoring examples. For instance, helping a family find a house scored 100 points. The AI found properties that met budget and commute requirements, calculated property transaction taxes, and added relevant documents to the shopping cart for user review. Choosing a kindergarten after moving scored 93.1 points, with the AI screening three candidate schools and compiling enrollment information.

LifeVal has previously been used to guide model improvements and currently remains an internal benchmark designed and judged by OpenAI itself.

Original link https://m.theblockbeats.info/flash/375200

Key points

Dongcha Beating AI stated that OpenAI released a new evaluation suite called LifeVal to test AI's ability to handle life affairs.
The report said that OpenAI drew on survey data about the daily lives and time allocation of residents in various countries, and asked 3995 people what they most wanted AI to help with.
LifeVal covers areas such as parenting, healthcare, shopping, financial management, and education.
The average scores listed in the report are: GPT-5.2 at 49.3%, GPT-5.4 at 73%, GPT-5.6 at 80%, and GPT-6 Astra at 80.6%.
The report showed that the apartment-hunting case scored 100 points, and the kindergarten selection case scored 93.1 points.
LifeVal is currently still an internal test designed and evaluated by OpenAI itself.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

LifeVal uses tasks such as finding a kindergarten, buying groceries, applying for government benefits, and caring for sick family members to check whether AI can handle real-life matters step by step. The higher the score in the report, the better the model performs in this internal test, but it cannot directly represent its effectiveness in all real-world scenarios.

Why it matters to readers

This evaluation focuses on the concrete life tasks of ordinary people, rather than just answering knowledge questions; the disclosed results also show that different models perform differently on this test.

Beginners can understand it as an "AI life task capability exam." When selecting or using relevant models, one should still review the results found by the AI, such as housing listings, schools, fees, and documents.

Risks and unknowns

The evidence does not include official OpenAI announcements, papers, evaluation sets, or method descriptions, making it impossible to independently confirm the release status and scores.
LifeVal was designed and evaluated by OpenAI itself, with no evidence proving that it has been reproduced by third parties.
The evidence does not specify the scoring criteria, the scale of test tasks, or model availability.
High-scoring cases do not prove that the model can achieve the same performance in other life tasks.

OpenAI reportedly releases LifeVal life-task benchmark, disclosing scores across multiple generations of models

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center