OpenAI launches LifeVal, scoring its agents on everyday life tasks

Summary

• OpenAI published LifeVal on November 2, 2026, an internal benchmark grading agents on daily-life tasks.
• Scores rose from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra.
• The task set, grading rules and leaderboard are unpublished, and only OpenAI models were tested.

The latest

OpenAI's newest model clears roughly four in five everyday errands in the company's own grading, according to LifeVal, a benchmark the company published on November 2, 2026. It scores models acting as agents on tasks such as childcare, health, personal finance and dealing with government services — not coding or exam questions. GPT-6 Astra scored 80.6%.

Details

• The method: Models run the tasks as agents using web browsing, computer use and connectors, OpenAI said. Seven categories are covered: caring for others, health, education, running a household, personal finance, shopping, and accessing government services. The headline metric is average reward, unweighted.
• The scores: At xhigh effort, OpenAI reported GPT-5.2 at 49.3% on December 11, 2025, GPT-5.4 at 73.0% on March 5, 2026, GPT-5.5 at 77.4% on April 23, 2026, GPT-5.6 at 80.0% on July 9, 2026, and GPT-6 Astra at 80.6% on September 3, 2026.
• The curve: The company's own figures show the bulk of the gain landing between December 2025 and March 2026, a jump of 23.7 percentage points. The three releases after April 2026 added 3.6 points combined, with the latest model improving 0.6 points on its predecessor.
• Where tasks came from: OpenAI said the task list draws on the American Time Use Survey and the Multinational Time Use Study, plus a company survey of 3,995 participants it describes as demographically representative. Respondents were asked where they would want help from a personal assistant.
• Sample results: Among tasks OpenAI selected to showcase: finding a family a new home scored 100%, coordinating volunteers ahead of a flood 98.0%, watching a match remotely with a grandmother 96.3%, enrolling a child in kindergarten after a move 93.1%, and preparing meals for a parent newly diagnosed with diabetes 90.4%.
• The company's framing: OpenAI said that to be a truly helpful personal assistant, "ChatGPT should be able to take actions that actually get things done." The framing positions the benchmark as a measure of delegation rather than knowledge.
• What is withheld: The task set and the grading rubric are not published, there is no public leaderboard, and the evaluation covers OpenAI models only. The company did not say whether any of those will be released.
• The comparison gap: Because no rival systems were run and no scoring rules were released, the figures cannot be checked against competing models or reproduced by outside researchers. OpenAI presented the results as internal evaluation.

Background

Most widely cited model benchmarks measure coding, mathematics or exam-style reasoning. LifeVal is pitched instead at agentic execution — whether a model can complete an errand end to end using browsing, computer use and connected services, rather than answer a question about it.

Between the lines

The showcased examples all cluster above 90%, while the overall average sits at 80.6% — a gap implying a tail of tasks performing well below the published figures, which the categories alone do not identify. The flattening since April 2026, 3.6 points across three releases, suggests the benchmark may be nearing its ceiling as a way to separate successive models.

What's next

Watch for whether OpenAI releases the task set, rubric or a public leaderboard, whether rival labs are scored against LifeVal, and whether the next model release moves the average past 80.6%.

Source: OpenAI