A test built from how people spend their days
Most model evaluations start from what models find hard: exams, code, mathematics. LifeVal starts from the other end. OpenAI says it took two studies that record how people actually spend their time, the American Time Use Survey and the Multinational Time Use Study, and added a survey of 3,995 respondents, described as demographically representative, asking where people would most value help from a personal assistant. The tasks cover caregiving, health, education, household administration, personal finances, shopping and government services: finding childcare, preparing for school, making a budget, ordering groceries, dealing with official paperwork.
The framing is explicit about where OpenAI wants ChatGPT to go. A helpful assistant, the page argues, "should be able to take actions that actually get things done", not stop at advice. LifeVal is the yardstick for that move: OpenAI says it has used the evaluation for the past year to track and drive its models' performance on everyday tasks, training them on browser use, computer use and connectors along the way. OpenAI is optimizing for errands, and it has now published the scoreboard it has been using — most of it, anyway, and in a place a casual reader has to know to look: the headline numbers sit in the page's charts and an expandable results table, not in its prose.
The average: 49.3% to 80.6%
The chart the page leads with is a progression. Each point averages a model's attempts within each of 169 tasks, then averages those task means equally. The resulting metric is labeled "Average reward (unweighted)" and displayed as a percentage. GPT-5.2 recorded 49.3% in December 2025. GPT-5.4 reached 73.0% in March 2026, GPT-5.5 77.4% in April, GPT-5.6 80.0% in July, and GPT-6 Astra 80.6% in September, using the xhigh setting identified in the launch blog. There is no willingness-to-pay weighting; the chart expresses reward multiplied by 100 and rounds labels to one decimal place.
That is a real overall result, stated with its method attached, and it deserves to be reported as one. It also carries its limits on its face. The averages are OpenAI's, on OpenAI's tasks, marked by OpenAI's reward; the steps between models compress noticeably at the top of the chart, from a 23.7-point jump between GPT-5.2 and GPT-5.4 to 0.6 of a point between GPT-5.6 and GPT-6 Astra.
Separately, OpenAI's API comparison refers to GPT-6 rather than explicitly identifying the GPT-6 Astra xhigh configuration used for the first chart.
The chart says the newest model has a slightly higher average reward. OpenAI's API comparison adds a qualification: GPT-5.6 achieves similar overall performance on these tasks at roughly half GPT-6's estimated API cost, while GPT-6 is about 23% faster. For buyers, that is a more useful comparison than the small difference between the two models' headline averages.
The table: best attempts, ten of them perfect
Click "See more" and the page shows a "Task results" table whose single score column is headed "Best recorded GPT-6 Astra score". Sixteen tasks are listed. Ten sit at 100%: finding and booking a sensory class for a toddler, booking a car inspection before a road trip, ordering halal groceries within a budget and itemizing them in a spreadsheet, replacing a lost engagement ring from a photo and a diamond report, picking one pocket-sized device to replace a phone, an e-reader and a handheld console, unsubscribing from email lists, choosing and buying a camera for a parent, finding a theater workshop a child can get to, booking two seats at an April piano concert in Ann Arbor, and finding a new home that fits a budget. Below them: a Texas SNAP application at 98.7%, Geneva travel on credit card points at 96.7%, and the lowest row shown, reviewing LSAT Logical Reasoning answers, at 88.7%.
"Best recorded" is doing important work in that heading. These are the highest recorded scores OpenAI chose to show for 16 individual tasks, not the average performance on those tasks. The broader chart measures something different: average reward across all 169 task means, with each task weighted equally. Ten perfect scores in the showcase therefore do not mean GPT-6 Astra normally completes those tasks perfectly. Nor can the 80.6% overall average tell us the average score on these particular 16 tasks. The two views answer different questions, and OpenAI does not publish enough underlying results to connect them.
Both numbers are OpenAI's, and both are accurately labeled. They describe different things: an average across 169 tasks and the best recorded scores for 16 selected tasks. The page provides no per-task distribution or repeated-run results with which to measure consistency. The showcase demonstrates what the model has achieved on selected tasks; the average provides a broader measure of performance. Neither tells readers how reliably the model will handle the same task on the next attempt.
What OpenAI still does not publish
The methodology is more visible than the headline alone suggests, but the evaluation is not reproducible. OpenAI identifies the 169-task count and explains how the overall average is calculated. It also publishes five prose examples and sixteen best-recorded task results. What remains private is the complete task set, the underlying dataset, the full per-task results and the code. The page also does not explain the reward rubric in enough detail for an outside group to reproduce the scores or run a competitor against the same evaluation.
None of that makes LifeVal useless. An internal evaluation that a lab trains against for a year will shape its product, and the direction it reveals — toward acting on personal errands through browsers, computers and connected services — is real information about what ChatGPT is being built to do. The progression is also a genuine claim that can be watched: if the average stalls near 80% while the best-recorded table fills with 100%s, that divergence will say more than either number alone. LifeVal is a reliable guide to OpenAI's priorities and a partially checkable guide to its progress. It is not yet a scoreboard anyone else can use.
For a reader deciding whether to hand an assistant a real errand, the practical test remains the unglamorous one: give it a task of yours, with your constraints, and check the output before it touches anything that matters. An 80.6% average reward across OpenAI's 169 tasks, alongside best-recorded scores on sixteen selected tasks, cannot tell you how reliably an assistant will handle your own requirements. That still needs to be tested in practice.
Sources
- [1] OpenAI, "LifeVal", openai.com, November 2, 2026 (primary source; read first-hand in full on the fully rendered page in a live browser on November 3, 2026 — the prose; chart 1 "Model performance on everyday life tasks over time" and its description: 169 tasks, unweighted average reward, percent = reward multiplied by 100, no WTP weighting, GPT-6 Astra at the launch-blog xhigh setting, and the five dated model points 49.3% / 73.0% / 77.4% / 80.0% / 80.6%; the five prose examples under "Our current models perform well on many example tasks"; the "See more" dialog "Task results" with the column "Best recorded GPT-6 Astra score" and its 16 rows, scrolled to the bottom; the chart-2 prose on GPT-5.6's similar overall performance at roughly half the estimated API cost and GPT-6 being about 23% faster)Read source
- [2] AIWeekly, "OpenAI's LifeVal scores GPT-6 Astra 80.6% on everyday tasks", November 2026 (its 80.6% / 49.3% progression matches the rendered page's chart 1; cited as corroboration that the chart figures were the ones circulating)Read source
- [3] explainx.ai, "OpenAI LifeVal: Everyday-Task Agent Benchmark Explained", November 2026 (its "best recorded" table description matches the rendered dialog; its other secondary-only claims are not used)Read source