Beyond Answering Questions
For years, the promise of artificial intelligence has been a true digital assistant—a system that could not just answer questions, but take action. While today's AIs can write an email or settle a bar bet, they often struggle with the complex, personalised
tasks that define daily life. Ask one to find a reliable nanny with specific certifications, manage a cross-country move, or handle government paperwork, and the illusion of competence can quickly fade. This is the gap OpenAI is trying to close with LifeVal, a new evaluation framework introduced on November 2, 2026. Instead of testing for academic knowledge, LifeVal measures an AI's ability to be genuinely helpful by successfully completing the real-world jobs people want to delegate the most.
What Makes a Task 'Real'?
To build a benchmark that reflects genuine human needs, OpenAI didn't just guess what matters. Researchers combined data from large-scale time-use studies with a new survey of nearly 4,000 people to identify which tasks people would most want a personal assistant to handle. The result is a set of 148 task categories that represent the friction points of modern life: finding childcare, managing personal finances, planning a diet for a family member with health issues, and navigating bureaucracy. The tasks are designed to be 'agentic', meaning they require the AI to do more than talk; it must plan, act, and complete a workflow, often using personal context and files to get the job done right.
How LifeVal Measures Success
Unlike traditional benchmarks that might give a simple pass/fail grade, LifeVal uses a more nuanced approach. For each task, human experts with real-life experience in that specific area create a detailed rubric for what a successful outcome looks like. For example, when asked to help a parent find a new kindergarten after a move, the AI isn't just graded on whether it found a list of schools. The evaluation checks if it considered the user's budget, commute, educational preferences, and other personal details mentioned in past conversations. Success is not just providing an answer, but completing the task in a personalised and genuinely useful way. This shift from 'can it answer?' to 'did it help?' is a critical evolution in how AI capabilities are judged.
Early Results and Future Ambitions
OpenAI has been using LifeVal internally for the past year to drive improvements in its own models. The results show significant progress. According to the company, its latest model, GPT-6 Astra, achieved an overall score of 80.6%, a huge leap from the 49.3% scored by a late 2025 model. On specific tasks, the scores are even more impressive, with one model hitting 93.1% on selecting a kindergarten and 98% on managing community volunteers for flood preparation. However, these numbers currently only apply to OpenAI's own models, as the full benchmark has not yet been released publicly. The flattening of performance gains around the 80% mark also suggests there is still room for improvement, either in the models or the complexity of the tasks themselves.
