OpenAI Scored AI at 80.6% on Everyday Tasks. An Independent Benchmark Published the Same Day Scored 12% — Insurers Should Price the 12%

OpenAI published LifeVal on Nov 2, 2026 to measure whether an AI assistant finishes everyday errands rather than answering questions about them: find childcare, enroll a child in kindergarten, order groceries on a budget, reach government services. The models on its own chart run from 49.3% to 80.6% on \"average reward.\" The same morning, SemiAnalysis published GDPx, where the best model from any lab manages just 12.0% of 206 long-horizon tasks in 11 economically valuable domains. For an insurer or a benefits administrator, the second number describes the job.

Neither instrument answers the question a benefits administrator is actually asking. GDPx forgoes government, real estate and insurance by design; LifeVal's task set is unpublished, its grader undefined and its chart contains only OpenAI's own models. The pair leaves the regulated-service middle unmeasured — precisely the paperwork-heavy, permission-bound services where a wrong answer costs money and gets audited.

LifeVal's chart is a progress curve, and it has decelerated hard: 23.7 points from GPT-5.2 to GPT-5.4, then 4.4 points, then 2.6, then 0.6 points to GPT-6 Astra. Three releases bought 31.3 points; the fourth added only 0.6 points in its most recent release. The task areas are caregiving, health, education, household administration, personal finances, shopping and government services, and OpenAI's own worked examples land between 90.4% and 100% — finding a new home for a family scores a clean 100%.

A curve that flattens at 80% while an independent benchmark puts frontier agents at 12% on hard professional work says the constraint moved out of the model and into the world: accounts, permissions, forms and institutions. An agent can find childcare at a 100% score and still be unable to enroll the child, because enrollment needs someone else's authorization.

GDPx publishes what each model costs per task, and those prices are the reason service delivery is about to be repriced. GPT-6 Astra runs a long-horizon task for $53.99 on average; the cheapest model on the leaderboard does it for $0.94; the most expensive costs $67.16. Every task carries an eight-hour budget.

OpenAI has already put that price against a real service transaction. In its May 2026 pilot with Thrive Holdings across Crete's network of more than 30 accounting firms, a tax agent drafted 7,000 returns, saved practitioners about a third of their preparation time, lifted throughput roughly 50% and reached up to 97% drafting accuracy in six weeks — only about a quarter of returns met a 75% field-completion bar at launch, and 86% met it within six weeks. One senior accountant cut preparation from 180 hours last season to 15 hours this year.

The arithmetic is the whole story. The cost of thinking has fallen far enough that the binding constraint on service delivery is no longer the model's price per answer but who is authorized to act on the customer's behalf, and who carries the error when the agent files the wrong form.

Follow the gross margin. Vendors selling picks and shovels and owners of the record keep the economics while handling-hour vendors get squeezed. On each company's latest reported trailing twelve months, the ladder runs from 74.7% at NVIDIA and 66.5% at Thomson Reuters, through 47.8% at Lemonade and 47.1% at Tyler Technologies, down to 36.6% at Genpact and 18.8% at Alight.

Genpact sold $5,079.9 million in 2025 with an average headcount of about 146,000 — generating roughly $34,800 of revenue per employee, which is the arithmetic the agent attacks. Lemonade is the counter-example: it reports a claims automation yield above 50%, has grown in-force premium per employee about threefold in four years, and produced a record $113 million quarterly gross profit on $294 million of revenue while still running a $19 million adjusted EBITDA loss.

The clearest sign that this is a live shift rather than a forecast sits in a segment footnote. Intuit's FY2026 10-K shows TurboTax revenue up $363 million and says the increase came from assisted tax and consumer money offerings, \"partially offset by a decrease due to fewer TurboTax federal units.\" ProTax, which sells software to preparers, grew 4% to $647 million, and the whole company still grew revenue 14% to $21.4 billion. Intuit reported fewer TurboTax federal units inside a growing franchise.

Thomson Reuters is the trade from the other side: on Sept 17, 2026 it put CoCounsel Legal and HighQ behind a Model Context Protocol connection into ChatGPT Enterprise, so the record, the content and the permissioning move inside the assistant instead of standing in front of it. That is what owning the rails looks like once the assistant becomes the front door.

The next quarter puts the thesis to test on specific days, and each catalyst measures a different part of the mechanism.

Over one to three years the durable question is not whether an agent can file a tax return — OpenAI has already run that on a real season of work — but who is allowed to bind the institution. LifeVal's tasks were built from how people actually spend time, the American Time Use Survey and the Multinational Time Use Study plus a survey of 3,995 respondents, and its categories are the services that need somebody else's permission: school enrollment, government services, personal finances. ChatGPT now reaches more than a billion people every week, so the front door for these services is consolidating into one place.