Anyone who has dealt with government paperwork probably knows this: the exhausting part isn’t looking things up, it’s the step right before you hit submit. One missing document or one wrong field means doing the whole trip again, so most people re-check everything from scratch themselves.
That’s the first thing I thought of when I saw OpenAI announce LifeVal.
What This Test Is Actually Measuring
LifeVal is a new benchmark OpenAI announced on November 2, built to test whether AI can handle the everyday chores of ordinary people. Finding a preschool, buying groceries, applying for government benefits, caring for a sick family member: these are all things that have to be handled one step at a time.
I think the way they sourced the tasks was thoughtful. OpenAI looked at time-use surveys from the US and several other countries to see where people’s days actually go, then asked 3,995 people directly what they most wanted AI to help with. So the coverage is childcare, healthcare, shopping, personal finance, and education. These aren’t exam questions, they’re just life.
80.6, But the Curve Flattened Long Ago
On OpenAI’s chart, GPT-5.2 scores 49.3, GPT-5.4 jumps to 73.0, GPT-5.5 gets 77.4, GPT-5.6 gets 80.0, and the newest GPT-6 Astra gets 80.6.
I read that list of numbers twice. From GPT-5.2 to 5.4, the score rose almost 24 points. The three generations after that added only 7.6 points combined, and the latest generation is just 0.6 points above the previous one.
So the 80.6 in the headline is real, but the story it tells isn’t “AI is getting better and better at running a life.” It looks more like “it learned the easy parts fast, and now it’s stuck on a smaller, harder chunk.”
The 100-Point House Hunt: How Far Did It Actually Go?
OpenAI showed some high-scoring examples. According to BlockBeats’ write-up, in the task of helping a family find a home, which scored 100, the AI found listings that fit the budget and commute requirements, calculated the property transaction tax, added the relevant documents to a cart, and handed it to the user to review. In the task of picking a preschool after a move, which scored 93.1, the AI narrowed it down to three candidate schools and organized the enrollment information. (I didn’t get to read the official page directly this time, so the details of these examples rely on secondhand write-ups.)
What caught my eye was the verb: added to a cart, for the user to review.
What that 100 achieved is “getting everything ready so all you have to do is press the button.” And the most tiring part I mentioned at the start sits right before that press. The AI finishes 90% of the work, but you don’t dare trust it blindly, so you still re-check everything yourself, which means the time saved isn’t as much as the score makes it look. The test counts “preparation” in the score and leaves “peace of mind” outside it.
Writing Its Own Questions and Grading Its Own Answers
There’s one more thing I need to say plainly. LifeVal is an internal test that OpenAI designed and judges itself. It has already been used over the past year to guide model improvements, and the only scores published are for OpenAI’s own models.
How many tasks there are in total, how the average is calculated, whether there’s a fixed scoring rubric, and whether it runs on real websites or test replicas: none of that has been made public so far. Also, the official page lists six tasks where GPT-6 Astra scored 100 on its “best run,” and that is a best result, not an average.
Using your own questions to track your own progress is exactly what an internal test is for. But if you want to use 80.6 to judge whether “AI can handle my household stuff right now,” I’d pause first.
The Curve Flattened, and I Suspect It’s Stuck on the Last Step
The official write-up doesn’t say why the curve flattened, so what follows is my own guess. The things it learned quickly were probably the “preparation” kind of work: looking things up, comparing, organizing. What it’s stuck on is more likely the actions you can’t take back, plus all the exceptions in the real world that aren’t written in any rulebook. More rounds of training may not move the score on that kind of thing, because the hard part is no longer knowing the answer. And even if the test hands out a 100, the friction of the real world doesn’t go away.
If You’re Using AI to Get Things Done Yourself
When building automations, instead of chasing fully hands-off from start to finish, deliberately leave a “shopping cart.”
Let the AI finish the research, the calculations, and the drafts, and put all of it in a pending-confirmation area, but make sure the final send always stays in a human’s hands. The point of tuning isn’t to make it smarter, it’s to make review faster: attach a source link to every conclusion, have it flag the parts it’s unsure about, and then a person only needs to look at the flagged items.
If you’re also automating things for your own small business, try thinking about this one question first: in your workflow, which step can’t be undone once you press it? Before that step, it’s fine to let the AI go all the way to 100.
That’s what I look at when I look at LifeVal’s scores.