OpenAI Launches LifeVal Benchmark for Everyday Tasks, With GPT-6 Astra Scoring 80.6%

OpenAI released the benchmark on November 2, according to AI News Weekly. It covers seven areas of daily life: caregiving, health, education, household administration, personal finances, shopping and dealing with government services. Models work through the tasks as agents, using browsers, computer use and connectors to other services.

The company said it built the task set from the US Bureau of Labor Statistics' American Time Use Survey, the Multinational Time Use Study, a demographically representative survey of 3,995 people, and about 100 participants who recorded tasks they wanted an assistant to take on. Results are reported as an average reward.

OpenAI said a truly helpful personal assistant should be able to "take actions that actually get things done". It argued that a model tested in a chat-style evaluation "can complete the required clicks and still produce an unsuitable result".

Scores Across OpenAI Models

GPT-6 Astra scored 80.6 per cent, up from 49.3 per cent for GPT-5.2. OpenAI's chart puts GPT-5.4 at 73.0 per cent, GPT-5.5 at 77.4 per cent and GPT-5.6 at 80.0 per cent. Astra's best results included kindergarten enrolment after a family move, at 93.1 per cent, and managing volunteers ahead of a flood, at 98.0 per cent.

Caveats

The results come with limits. OpenAI built, ran and graded the benchmark itself, and every score is self-reported. The task set is not public, there is no leaderboard, and models from other companies cannot be tested. Reports also differ on its size: one developer write-up says it has 169 tasks, while AI News Weekly said OpenAI's page gave no count. The 0.6-point gain from GPT-5.6 to Astra is small.

For companies looking at consumer and employee AI agents, LifeVal shows OpenAI trying to measure agents by results rather than by how they reason. Until outsiders can check the tasks and the grading, though, the scores are vendor claims, not independent evidence.