OpenAI on November 2, 2026 published LifeVal, an internal evaluation of how its models handle everyday personal-assistant work. The post says the lab used LifeVal over the past year to track and drive gains on tasks people ask ChatGPT to help finish, not only to answer about.
The headline chart OpenAI shows on the page puts GPT-6 Astra at an 80.6% unweighted average reward and GPT-5.2 at 49.3%, matching independent write-ups that read the same chart. Selected example cards on the page score much higher; those are individual demos, not the average.
Key Facts
| Item | Detail |
|---|---|
| Name | LifeVal |
| Publisher | OpenAI (research post, Nov 2, 2026) |
| What it measures | Model performance on everyday personal-assistant tasks |
| Task areas named | Caregiving, health, education, household administration, personal finances, shopping, accessing government services |
| How tasks were chosen | American Time Use Survey, Multinational Time Use Study, plus a demographically representative survey of 3,995 respondents |
| Headline scores (OpenAI chart, unweighted average reward) | GPT-6 Astra 80.6%; GPT-5.2 49.3% |
| Intermediate points reported from the same chart | GPT-5.4 73.0%; GPT-5.5 77.4%; GPT-5.6 80.0% |
| Example cards on the page (selected demos) | Kindergarten enrollment after a move 93.1%; flood volunteer coordination 98.0%; finding a home 100%; remote game watch with a grandparent 96.3%; meal plan after a diabetes diagnosis 90.4% |
| Capabilities OpenAI ties to gains | Browser use, computer use, and connectors |
- Metric wording. OpenAI and secondary reports describe the chart figure as average reward, unweighted across tasks. The page does not publish a downloadable task set or grader.
- Scope. Charts and examples cover OpenAI models only. No third-party model scores appear on the post.
Background
OpenAI frames LifeVal against academic knowledge tests. The lab says it started from time-use surveys that catalog daily activities in the U.S. and internationally, then asked where people most want personal-assistant help. The resulting scenarios include finding childcare, navigating government services, budgeting, shopping, and related household work.
The company says it trained models over the past year to be faster and more capable on browser use, computer use, and connector workflows, and that enabling those skills in ChatGPT widens the set of tasks the product can attempt. LifeVal is presented as the yardstick used while that training ran.
Reactions and Impact
Coverage from outlets that inspected the post, including Brocker, AI Affairs, and Nerds Magazine, repeats the same 80.6% / 49.3% chart pair and notes the same limits: OpenAI grades its own models, the task set is not public, and the sample cards are high-scoring picks rather than the mean. Brocker also flags that GPT-6 Astra's 80.6% sits only 0.6 points above GPT-5.6's 80.0% on the published timeline.
What We Don't Know Yet
- Exact task count, rubrics, and grader calibration (OpenAI has not released them)
- Whether LifeVal will be opened for independent reruns or cross-vendor leaderboards
- Scores for models released after the chart's GPT-6 Astra point
- How average reward maps to success rates in uncontrolled real-world use