80.6% at Chores, 12% at Real Work: AI's Benchmark Week

Posted on Nov 5 • Originally published at toolpilotpro.com

On Monday, November 2, 2026, two benchmarks landed on the same day and told two opposite stories about the same machines. OpenAI published LifeVal, its new 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%. SemiAnalysis published GDPx, a test of real professional work — chip design, legal analysis, quant trading, eight hours per task — and the best model in the world, Anthropic's Claude Opus 5.5, completed 12.0% of it. Same week, same generation of AI: four-fifths of the chores, one-eighth of the expert work.

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 selling 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. That is what this guide works through, with both tests graded as honestly as they grade the models.

Most benchmarks test what AI researchers find interesting: exams, code puzzles, maths contests. LifeVal tests what fills an ordinary week. OpenAI built the task list from two time-use studies — the American Time Use Survey and the Multinational Time Use Study, which record how people actually spend their days — plus a new survey of 3,995 people asking where they would most want a personal assistant's help. The result is a set of tasks across caregiving, health, education, household administration, personal finances, shopping and accessing government services.

The tasks are agentic, and they ask the model to finish the job rather than describe it. The model browses, operates a computer and calls connectors, and a scenario-specific rubric grades the attempt. OpenAI's framing is explicit: to be a truly helpful personal assistant, ChatGPT should take actions that actually get things done.

Its published examples show the flavour — and the scores its current models achieve on them: enrolling a child in the right kindergarten after a move (93.1%), organising volunteers ahead of a flood (98.0%), finding a new home for a family (100%), planning how to watch a game with grandma (96.3%), and preparing meals 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.

The headline metric is average reward: each task's attempts are averaged, then the task averages are combined with equal weight, on a 0-to-1 scale. It is partial credit by design — a task done 80% well scores about 0.8. OpenAI's chart plots its own models at high reasoning effort against their release dates:

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.

LifeVal is a useful signal and a self-reported one, and OpenAI does not hide the arrangement: the company says it used the evaluation "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. The limits are structural:

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 — and the plateau it reveals is genuinely informative, because OpenAI had every incentive to show a steeper line. But it is a progress report, not a comparison. The model whose maker built LifeVal is, in our testing experience, a strong everyday assistant — see our ChatGPT review — yet nothing in LifeVal proves it beats the rivals it never faced.

GDPx comes from the opposite direction. SemiAnalysis — the semiconductor and compute research firm led by Dylan Patel, with the benchmark authored by Max Kan, Andrew Megalaa and Patel — built it as a successor to OpenAI's earlier GDPval, which the authors call "woefully inadequate for evaluating today's frontier models." 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, chosen for economic return on investment rather than their share of GDP: chip design, hardware design, materials science, cybersecurity, biology, alignment, finance, legal, quant trading, recursive self-improvement and software engineering. Government, real estate and insurance are deliberately skipped. The tasks were built with 14 specialist data partners — firms such as Arithmetic (cybersecurity), Latch Bio (biology), Chakra (legal), Third Moment (quant trading) and Datacurve (software engineering) — with SemiAnalysis running final quality checks.

Each task runs inside a simulated workplace for up to eight hours, in the model's own harness: Claude Code for Anthropic's models, Codex for OpenAI's, Antigravity for Google's, OpenCode where a model has none. The headline score is pass@1 — the first-attempt pass rate — averaged over three attempts, with every domain weighted equally regardless of how many tasks it holds.

Thirteen models were ranked; the six above are the ones whose scores launch coverage consistently reports. Even the best agent fails roughly seven times out of eight. The detail underneath is spikier than the ranking suggests: Opus 5.5 wins quant trading, hardware design and chip design outright; Gemini 4 Carbon leads biology and, narrowly, software engineering; GPT-6 Astra wins recursive self-improvement and materials science.

And then there are the two zeroes. In finance and legal, under GDPx's intended pass-or-fail grading, basically every model scores exactly 0 — 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 Carbon is not yet generally available. The open-weight showing is a reality check for the local-models crowd: GLM 5.3, the best open model on the board, manages 2.7%, a reminder that the newest open releases we cover, like LongCat-2.5, are priced for agent endurance rather than expert depth. For the two closed leaders, our head-to-head in ChatGPT vs Claude vs Gemini reaches a similar pecking order on everyday work — GDPx suggests the order holds, barely, when the work gets professional.

GDPx aims its honesty demands at the industry, so it deserves the same treatment. Four caveats, all stated by SemiAnalysis itself or visible in its method:

None of that disqualifies GDPx either. A hard, economically chosen task set with published trajectories is a serious instrument, and SemiAnalysis plans updates every three to six months, which turns the 12% into a baseline the industry will be measured against. The benchmark's own best defence is that its numbers embarrass everyone equally — including the firmament's marketing departments. Anthropic's win here is real but narrow, and it comes from the model we rate for exactly this kind of long, careful work in our Claude review.

There is a second development inside this one, and it explains why two heavyweight tests arrived on the same Monday. Benchmarks used to be academic artefacts — published, downloadable, argued over in the open. LifeVal and GDPx are both, in different ways, commercial instruments.

OpenAI keeps LifeVal closed and uses it as an internal steering metric that doubles as a launch story; the evaluation "drove improvement" for a year before the public saw a single number. SemiAnalysis sells GDPx as a product to an industry that needs a frontier yardstick and cannot build one itself, with partner firms supplying the tasks and a leaderboard as the shop window. Even the older standard both are replacing, GDPval, was an OpenAI artefact.

The pattern generalises: as models converge, the power to define the test becomes a competitive asset, and the test-maker's incentives — sell subscriptions, sell a product story, sell the benchmark itself — sit inside every decimal point. That does not make the numbers useless. It makes the method section the most important part of any benchmark announcement, and it is why this guide spent as many words on rubrics, weighting and access as on the scores themselves.

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

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, 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%).

There is also a quieter agreement between the two tests, and it may be the week's most useful finding. LifeVal says everyday competence has flattened near 80% across two model generations. GDPx says expert competence starts near 10%. The distance between "the AI does my admin" and "the AI does my job" is exactly the distance between those numbers — and both benchmarks, from opposite ends, say that distance is still large. If your work is the admin kind, our guide to the best AI coding assistants and agent tools will serve you better than either leaderboard; if your work is the expert kind, GDPx is the first ruler that measures it in units an employer recognises.

It is self-graded, which is a weaker and fairer charge. OpenAI built the tasks, ran its own models and published only its own scores, with no public task set or rubric — so the 80.6% cannot be checked or compared. But the chart also shows an unflattering plateau (0.6 points over the previous generation), which a purely rigged test would not need to admit. Treat it as an honest internal progress report, not a competition result.

Because the test was built that way. GDPx's 206 tasks are long (up to eight hours), professional and graded mostly pass-or-fail, across domains chosen for difficulty and economic value; two domains are so hard that every model scores zero under strict grading. A low ceiling is the design goal of a frontier benchmark — it leaves room to measure progress. SemiAnalysis plans updates every three to six months, so the number to watch is how fast 12% climbs.

Trust each for what it measures. LifeVal is the better guide to consumer assistant work — with the caveat that only OpenAI models have taken it. GDPx is the better guide to professional agent work — with the caveat that it is a paid, largely private product whose tasks outsiders cannot audit. For a purchasing decision, weight the one whose tasks resemble your work, and read the independent coverage of both, such as the CellCog analysis of LifeVal and the GDPx launch summary, before either maker's own page.

It means that across LifeVal's tasks, the model's attempts earned on average 80.6% of the available rubric credit — partial progress counts. It does not mean the model fully completed 80.6% of the tasks. A run that books the right restaurant on the wrong day might earn most of the credit on LifeVal and a flat zero on a pass/fail test, which is most of why the two headlines differ.

Not today. OpenAI has not released the LifeVal task set, rubrics or graders, so the test cannot be reproduced outside OpenAI. GDPx is a commercial SemiAnalysis product whose tasks are largely private, though its leaderboard and run trajectories are public. The FourWeekMBA breakdown of LifeVal's arithmetic shows what outside analysts can and cannot reconstruct from the published charts.

November 2 gave the AI industry two mirrors. LifeVal shows a technology that has largely solved the shape of everyday assistance and is now improving at the margins — 80.6%, self-awarded, with the last two generations separated by less than a point. GDPx shows the same technology at the start of professional competence — 12.0% for the best, zero for everyone in finance and legal, and a leader that wins by spending three times more per task. Neither mirror is lying; they are angled at different work. The mistake is reading either headline alone. Read together, they draw the most honest map of late 2026 we have: AI that can run your errands is here, AI that can hold your job is not, and the benchmarks that will matter next are the ones built, like GDPx, to be failed.