OpenAI's LifeVal benchmark shifts the test from answers to finished tasks

OpenAI has released a benchmark called LifeVal that tests whether AI can complete everyday tasks rather than answer exam questions. The shift matters more to enterprises than the rumoured GPT-6.2 release, because it moves the question from "how smart is the model" to "can it finish the job".

The release that got buried

The loudest AI story of the past two days is GPT-6.2. According to IT Home, citing The Information, OpenAI may as soon as Friday local time (6 November) release a new generation of Astra and Sol series models, plus a Cyber model aimed at cybersecurity. The report gives no model numbers, pricing, context window or benchmark scores, and does not say whether these models will arrive first through the API, ChatGPT or Codex.

Meanwhile, something else went largely unnoticed: OpenAI published a benchmark called LifeVal. According to Xinzhiyuan, it does not test problem-solving but "getting things done" — everyday tasks such as finding a nanny, filing a tax refund or planning a cross-border move. The paper is hosted on OpenAI's website.

Our view: model launches are becoming routine. What enterprises should actually watch is the shift in how AI is evaluated — from "scores well" to "gets it done".

Why a benchmark matters more than a new model

According to Xinzhiyuan, the LifeVal tasks were not chosen at random. OpenAI examined the US Bureau of Labor Statistics' American Time Use Survey and time-use studies from multiple countries, covering how people in 25 countries spend their days. It then ran a willingness-to-pay survey of more than 3,800 real consumers matching the US census distribution, and finally selected 148 life tasks. The report says that if a perfect super-assistant completed these tasks, US consumers would collectively be willing to pay $1.7 trillion a year.

We cannot vouch for how that figure was calculated or whether it holds up. But the direction it points in is clear: the object of evaluation has moved from "knowledge" to "task closure" — seeing a task through from start to finish.

For the past few years, large-model vendors have competed on bar exams, maths olympiads and coding contests. Those scores mean little to enterprise buyers — you are not buying AI to sit exams, you are buying it to take one specific job from beginning to end. LifeVal writes "high context, personalisation, strong agency" into its test characteristics, which amounts to an admission: the hard part is not giving the right answer, but remembering your situation, completing the task on your terms, and leaving something that can be checked.

What does this mean for enterprises? It means that in future procurement, "high benchmark score" will explain less. The question to ask is: in my scenario, can it complete a task end to end, and can I audit what it produced?

The counter-argument: another marketing concept?

The most common objection is that the benchmark is published by the vendor, with tasks it chose and standards it set — so how is it different from self-praise?

The objection is valid, and partly valid. LifeVal was released by OpenAI itself, and the 148 tasks and the $1.7 trillion estimate all carry a vendor's perspective. It does not prove that OpenAI's models are the strongest.

But "self-defined standards" and "useless standards" are two different things. The real question is not who set the standard but whether outsiders can check it. According to Xinzhiyuan, the LifeVal paper is public, with the task list and survey methodology written into it; others can run the same tasks against their own models, or question whether the tasks are skewed. If it can be checked, there is room for discussion; only what cannot be checked deserves to be called marketing.

Note that even the new models themselves remain at the stage of subjective experience. At a moment like this, a public evaluation framework is scarce.

When this judgment would not hold

If LifeVal ends up as nothing more than a set of tasks in a paper — with no third-party replication and no vendor using it for side-by-side comparison — then it is a public-relations move, and our argument above should be discounted.

Conversely, if in the coming months other vendors start talking in terms of similar task-based benchmarks, and enterprise buyers start asking "how does it perform on my business tasks", then the shift is real.

What is certain now: the narrative around model capability is moving from "smarter" to "more able to get things done". What is uncertain is whether the industry will pick up the new ruler.

One thing enterprises can do now

Do not wait for GPT-6.2. First write down your own "LifeVal" — the tasks in your business that consume the most people, go wrong most often and most need item-by-item checking. Spell out what the input is, what the output should look like, and how errors are caught.

This is more useful than chasing new models, because models will keep changing while your task list will not.

As it happens, SHEYU ZHISHEN (舍予AI智审) addresses one slice of this kind of work: it turns technology project application materials into a checkable evidence table, reading tens to hundreds of pages in seconds. Every extracted value carries a page number and the original text, and conclusions can be traced back to the original guideline. Its four-value results do not pass by default, and data does not leave the internal network. It handles only application materials — other scenarios are beyond it. What you need is to work out what your own tasks look like first, then find the tool that matches them.