OpenAI Releases LifeVal: Agents Tested on Everyday Errands, Example Scores Top Out at 100%
OpenAI's official page introduces LifeVal, whose task list derives from time-use surveys and a questionnaire of 3,995 respondents. The example scores are self-reported by OpenAI, cover only its own models, and are not yet open to independent reproduction.
- 01THE ANNOUNCEMENT — OpenAI's official page published LifeVal on November 2, releasing scores for five everyday-task examples.
- 02THE CONFIRMATION — Multiple independent reports relayed the page's chart; the task sources and example scores were cross-checked for consistency.
- 03THE WATCHPOINT — Whether the task set and scoring rubric are published, and whether third parties can independently reproduce the results.
WHAT HAPPENEDWhat Happened
OpenAI's official page says LifeVal measures a model's ability to complete everyday tasks across seven domains — caregiving, health, education, household administration, personal finance, shopping and public services — with models acting through browser use, computer use and connectors to get things done, not merely give advice.
The source of the task list is documented on the official page: US time-use surveys and cross-country time-use research provide the catalog of everyday tasks, supplemented by a demographically representative questionnaire of 3,995 respondents asking which tasks people would most like to hand to a personal assistant.
The five example scores published on the official page: choosing a suitable kindergarten after a move, 93.1%; organizing volunteers ahead of a flood, 98.0%; finding a new home for a family, 100%; accompanying a grandmother to a game remotely, 96.3%; and a meal plan for a parent newly diagnosed with diabetes, 90.4%. The page says the latest model scores highly on many LifeVal tasks.
Multiple independent reports relayed chart data from the page showing GPT-6 Astra averaging 80.6% on LifeVal, up from GPT-5.2's 49.3% in December 2025. That chart figure comes from reporting; the official page's main text does not publish an overall score.
OpenAI says it has used this evaluation over the past year to track and drive model performance on everyday tasks, with the aim of moving ChatGPT from a chat tool toward a personal assistant that genuinely gets things done.
WHY IT MATTERSThe exam questions grow out of real time allocation
LifeVal's task list was not invented by researchers. It starts from where people actually spend their time, then asks what they would most like to hand off. Finding childcare, running government errands, ordering food and cooking — these are exactly where ordinary people spend real hours every day.
That positioning pushes the agent battleground from the office into the home and consumer services. The trio of browser use, computer use and connectors decides whether an assistant can move beyond giving advice and actually get things done.
WHY IT MATTERSThe scores are self-reported; the task list is the reusable part
LifeVal currently scores only OpenAI's own models. No task set, scoring code or leaderboard has been published, and the examples are high-scoring cases chosen by OpenAI, so outsiders cannot independently rerun them. Reading example scores above 90% as proof of real-world reliability would overstate what these numbers show.
What peers can genuinely borrow is the task-selection method: start from time-use data, then narrow with user surveys to the tasks people most want to outsource. Only when third parties reproduce the benchmark or the scoring rubric is published will LifeVal's scores support cross-model comparison.
SIGNAL JUDGMENT — ANALYSISWatch the task list, not the example scores
Tokenta Signal's judgment: the signal in LifeVal is not the example scores above 90%, but OpenAI publicly shifting its training focus toward everyday errands. The scores are the company's own; the task list is real. The next signal is whether third parties can reproduce it.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Confirmed by two sources: OpenAI's official page and independent reporting. Original link opened and verified.
PRIMARY SOURCESPrimary Sources
- OpenAI official page
https://openai.com/index/lifeval - ExplainX independent analysis
https://explainx.ai/blog/openai-lifeval-everyday-tasks-benchmark-gpt-6-astra-childcare-government-paperwork-2026
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook