OpenAI Says Its AI Scores 80% on Everyday Life — but Only 12% on Real Work. Both Numbers Are True

On Monday 2 November 2026, two benchmarks landed on the same day and told two opposite stories about the same generation of AI. OpenAI published LifeVal, a test of everyday life — enrolling a child in school, ordering groceries on a budget, wrestling a government office — and its newest model, GPT-6 Astra, scored 80.6%. On the same day, SemiAnalysis published GDPx, a test of real professional work — chip design, legal analysis, quant trading, up to eight hours per task — and the best model in the world completed 12.0% of it. Same week, same machines: four-fifths of the chores, one-eighth of the expert work.

Quick answer: OpenAI's LifeVal test puts GPT-6 Astra at 80.6% on everyday tasks like childcare and groceries — yet the best AI completes only 12% of real work. Both are true.

Both numbers are real, and neither means what a headline makes it sound like. The 80.6% is an average reward on a test OpenAI built, graded by OpenAI, on OpenAI models only. The 12.0% is a pass rate on deliberately brutal tasks, measured by a research firm that sells the benchmark. The interesting story is not which number is “true” — it is why the gap exists, what each test actually measures, and which one resembles the work you would hand an AI.

What is LifeVal?

LifeVal is OpenAI’s new evaluation for the work people actually ask a personal assistant to do. Where most benchmarks test what AI researchers find interesting — exams, code puzzles, maths contests — LifeVal tests what fills an ordinary week. OpenAI published it on 2 November 2026, and followed it with a full paper on arXiv on 4 November 2026 (arXiv:2611.06687, by Jenny Yang and colleagues).

The task list was built from three sources. OpenAI started with two time-use studies — the American Time Use Survey, which records how people in the United States spend their days, and the Multinational Time Use Study, the same kind of record across countries. It then ran a new survey of 3,995 people, which it describes as demographically representative, asking where they would most want a personal assistant’s help. The resulting tasks cover seven areas: caregiving, health, education, household administration, personal finances, shopping and government services.

The tasks are agentic, and that is the point. LifeVal does not ask the model to describe how to do something; it asks the model to finish the job, using browser use, computer use and connectors, with a scenario-specific rubric grading the attempt. The paper says OpenAI built 169 tasks across 148 consumer task categories, and that those categories represent more than $1.7 trillion in annual willingness to pay, estimated from a survey of 3,837 US consumers who were asked what they would pay for help from a “hypothetical perfect digital superassistant”.

“To be a truly helpful personal assistant, ChatGPT should be able to take actions that actually get things done.” — OpenAI, LifeVal, 2 November 2026

OpenAI is open about one thing the announcement does not hide: it used LifeVal “to track and drive improvement in model performance on everyday tasks over the past year”. In plain terms, the test was the training target before it was the press release. That does not make it useless — but it is the first thing to know before reading the scores.

The LifeVal scoreboard: 80.6%, and a plateau

OpenAI’s chart plots one number per model — an unweighted average reward across LifeVal tasks — against each model’s release date. Average reward is partial credit by design: a task done 80% well scores about 0.8, and the task averages are then combined with equal weight on a 0-to-1 scale.

Model Release LifeVal average reward
GPT-5.2 December 2025 49.3%
GPT-5.4 March 2026 73.0%
GPT-5.5 April 2026 77.4%
GPT-5.6 July 2026 80.0%
GPT-6 Astra September 2026 80.6%

Two things stand out. First, the shape of the progress: of the 31.3 points gained from GPT-5.2 to GPT-6 Astra, 23.7 arrived in a single step, between GPT-5.2 and GPT-5.4. The three releases since added 7.6 points combined, and Astra sits just 0.6 points above GPT-5.6. Everyday-task competence, on OpenAI’s own ruler, has plateaued at around 80% for two generations.

Second, the effort detail. A second chart breaks scores down by reasoning effort, and on GPT-6 Astra the “high” setting scores 80.7% while “extra high” scores 80.6% — the extra thinking costs more time and tokens for a fractionally lower score, with a “max” setting reaching 82.5%. More compute is no longer buying more chore competence. That plateau is arguably the most informative thing LifeVal reveals, because OpenAI had every incentive to show a steeper line.

What 80.6% looks like in practice

OpenAI shows five example tasks with scores. These are its chosen examples, not a random sample — but they show the flavour of the test, and the scores its current models achieve on them.

Example task Score shown
Enrol a child in the right kindergarten after a move 93.1%
Organise volunteers ahead of a flood 98.0%
Find a new home for a family, within budget 100%
Set up a way to watch a game remotely with grandma 96.3%
Prepare a first-week meal plan for a parent with a new diabetes diagnosis 90.4%

A separate list names six tasks where GPT-6 Astra’s best recorded attempt scored 100%, including booking a car inspection before a road trip, ordering halal groceries within a budget and itemising the order in a spreadsheet, and unsubscribing from email lists. Best attempts, note — not averages, and not typical attempts. The five example cards score 90.4% to 100%; the 169-task average for the best model is 80.6%.

The other number: 12% on real work

GDPx, published by SemiAnalysis on the same day, comes from the opposite direction. Where LifeVal asks whether an AI can run your errands, GDPx asks whether it can hold a job. The suite holds 206 tasks across 11 domains — chip design, hardware design, materials science, cybersecurity, biology, alignment, finance, legal, quant trading, recursive self-improvement and software engineering — chosen for economic return rather than their share of GDP. Each task runs inside a simulated workplace for up to eight hours. The headline score is pass@1, the first-attempt pass rate, averaged over three attempts, with every domain weighted equally.

Model Maker GDPx score Average cost per task
Claude Opus 5.5 Anthropic 12.0% $51.41
Gemini 4 Carbon Google 10.9% $16.29
GPT-6 Astra OpenAI 9.6% Not published
Muse Wonder Full Meta 9.0% Not published
Grok 4.7 xAI 3.9% Not published
GLM 5.3 Z.ai 2.7% Not published

Thirteen models were ranked; even the best agent fails roughly seven times out of eight. And the detail underneath is spikier than the ranking suggests. In finance and legal, under GDPx’s intended pass-or-fail grading, basically every model scores exactly zero — the work is either completed end to end or it is not, and it never is. SemiAnalysis therefore publishes average partial scores for those two domains instead, a quieter metric change readers should notice: the headline 12.0% blends binary passes in nine domains with partial credit in two.

Cost is the second story. Opus 5.5’s lead is bought, not just earned: it averages $51.41 per task and burns about 548,000 output tokens per task, against Gemini 4 Carbon’s $16.29 per task. And the pattern is not a one-off. Google’s own PLE evaluation, which we covered alongside BioBasicsBench, puts every agent under 50% on real engineering tasks. The chore score and the work score are measuring different worlds.

Why both numbers are true

Put the two rulers side by side and the “contradiction” dissolves. LifeVal and GDPx differ on every axis that determines a score.

LifeVal GDPx
Built by OpenAI SemiAnalysis
Measures Everyday chores and errands Expert professional work
Tasks 169 (per the paper) 206
Time per task Minutes-scale errands Up to 8 hours
Scoring Average reward (partial credit) Pass@1 (mostly pass/fail)
Models charted OpenAI only 13 models, all major labs
Task set public? No Mostly private (paid product)
Headline GPT-6 Astra 80.6% Claude Opus 5.5 12.0%

Three of those differences do most of the work. Partial credit versus pass/fail: a grocery order with one wrong item earns most of its LifeVal reward and would earn nothing under binary grading; an almost-finished chip design earns a GDPx zero. Task difficulty chosen on purpose: LifeVal samples the middle of ordinary life, while GDPx samples work selected because it is valuable and hard — its authors designed the ceiling so the best model “just” reaches 12%. Who is being measured: LifeVal’s chart is one family’s progress over time, GDPx’s is a cross-vendor snapshot — which is why Astra can be the hero of one (80.6%, its maker’s own ruler) and third on the other (9.6%).

What LifeVal cannot prove

None of this makes LifeVal worthless. A vendor-run progress chart, honestly labelled, tells you the direction and rough speed of improvement on the work consumers actually ask for. But four limits decide how much weight the headline deserves.

One vendor’s models. The chart contains only OpenAI systems. Claude, Gemini, Grok and every open-weight model are absent, and OpenAI has given no timeline for outside entries. Nothing in LifeVal proves Astra beats rivals it never faced.

Nothing is published. There is no downloadable task set, no public leaderboard and no published rubric or grader calibration, so nobody can rerun the test or check a score. OpenAI’s own post does not even state the number of tasks — the 169 figure comes from the paper, not the announcement.

Average reward is not reliability. A model can average 0.8 while failing the tasks that matter to you outright. The metric rewards partial progress, which is kind on chores and meaningless for tasks that are only worth doing completely — a tax filing done 80% well is a filing done wrong.

Selected examples. The task cards OpenAI shows are chosen highlights, and the 100% list is best attempts — marketing surfaces, not samples. The same caution applies to GDPx, which is a paid product with mostly private, partner-built tasks that no outside lab has audited, and whose scores are partly a measurement of how much each model’s safety classifier lets it attempt. Both tests deserve the same scepticism, applied in both directions.

Our take: read the ruler before the number

The most useful thing LifeVal gives you is not the 80.6% — it is a five-question test you can run on any benchmark announcement from now on.

  1. Who built it, and who does it grade? A test its maker also sells is a progress report, not a comparison. LifeVal grades OpenAI’s models; GDPx grades thirteen, but sells the test.
  2. Partial credit or pass/fail? The same work can score 80% on one ruler and 0% on the other. Ask which ruler matches the job — a grocery run tolerates partial credit, a contract does not.
  3. Can anyone rerun it? If the task set and rubric are private, the number is a claim, not a finding. LifeVal cannot be checked today; neither can most of GDPx.
  4. What does the trend line say, not the headline? LifeVal’s real story is the plateau: 23.7 of its 31.3 points arrived in one step, and the last three releases added 7.6 combined. Chores, on OpenAI’s own ruler, are nearly as solved as they are going to get this way.
  5. What would change the verdict? For LifeVal: an independent rerun, rival models on the same tasks, and a published rubric. For GDPx: a second lab reproducing the 12%, and the finance and legal domains scored the way they were designed to be scored.

Until those arrive, the honest summary is this: AI is now very good at the middle of ordinary life, on the test its maker built for exactly that, and still a long way from holding a professional job, on the hardest test anyone has built for that. Both numbers are true. The mistake is reading either one as the whole story. That is the same discipline we applied when reading METR’s acceleration assessment of Anthropic — and the same gap between a company’s own determination and an outside measurement that Anthropic’s own safety-threshold filing exposed this week.

Frequently asked questions

What is OpenAI LifeVal?

LifeVal is OpenAI’s benchmark for everyday agentic tasks — the chores and errands people want a personal assistant to finish, from finding childcare to ordering groceries and dealing with government paperwork. It was published on 2 November 2026, with a full paper on arXiv on 4 November. It covers 169 tasks across seven areas of daily life, built from two time-use studies and a survey of 3,995 people.

Does 80.6% mean GPT-6 Astra completes 80.6% of everyday tasks?

No. It is an average reward, not a completion rate. Each task is graded by a rubric that gives partial credit, so a task done 80% well scores about 0.8. The figure also covers only OpenAI’s own models, on a test OpenAI built and graded, with no public task set for anyone to check.

Why does GPT-6 Astra score only 9.6% on GDPx if it scores 80.6% on LifeVal?

Because the tests measure different things with different rulers. GDPx gives mostly pass-or-fail marks on up to eight-hour professional tasks in eleven domains, and Astra is third there behind Claude Opus 5.5 (12.0%) and Gemini 4 Carbon (10.9%). LifeVal gives partial credit on minutes-long everyday errands, graded by OpenAI itself. The gap is the yardsticks, not a contradiction.

Can I try LifeVal myself?

Not yet. OpenAI has published the announcement and the paper, but no task set, code or grader is public, and there is no public leaderboard. Until those exist, LifeVal is a vendor progress report rather than a test you can run or verify.

Does LifeVal mean AI can replace a personal assistant?

Not on this evidence. LifeVal shows OpenAI’s models score highly on the everyday tasks OpenAI chose, graded by OpenAI’s own rubric — and even there, the score has plateaued around 80% for two model generations. For tasks where partial credit is not enough, like taxes or contracts, the honest reading is the GDPx side of the story: the best model in the world still fails about seven times out of eight.

Sources and methodology

Claims in this article rest on OpenAI’s own LifeVal announcement (2 November 2026), the LifeVal paper on arXiv (arXiv:2611.06687, submitted 4 November 2026), and independent secondary analysis of both benchmarks. Every figure is labelled with its source; figures OpenAI or SemiAnalysis report about their own products are presented as their claims, not as independently verified facts. This article is desk-researched — the benchmarks were not run hands-on, and LifeVal’s task set is not public, so no independent reproduction was possible. Verification date: 15 November 2026.

Leave a Reply

Your email address will not be published. Required fields are marked *