OpenAI's LifeVal lacks a scoring method, and AWS's small agent lacks a bill

Two pages posted on November 2 give numbers that are easy to nod along to. OpenAI scores between 90.4% and 100% on five everyday tasks, from choosing a preschool to planning meals for a relative newly diagnosed with diabetes. AWS says a 4-billion-parameter model (parameters are the numbers inside a model; the more it has, the larger the model) matches the larger model that taught it on shopping tasks. Both figures were measured by the party presenting them, and each leaves out exactly what you need to decide: OpenAI’s page does not describe how the scoring works, and AWS’s post does not publish a cost per completed task.

OpenAI’s LifeVal: high scores on everyday tasks, scoring method not shown

On November 2, OpenAI posted LifeVal, an evaluation (a fixed set of tasks used to score a model) of how well AI handles daily chores: finding childcare, filing public-service paperwork, budgeting, preparing for school, ordering food. OpenAI says it chose the tasks from two studies of how people spend their time, the American Time Use Survey and the Multinational Time Use Study, plus a new survey of 3,995 respondents, and used the set to track model progress over the past year.

The page shows five examples with scores: choosing a preschool after a move 93.1%, coordinating volunteers ahead of a flood 98.0%, finding a home that fits a budget and a school-run commute 100%, setting up a way to watch a game remotely with a grandmother 96.3%, and a meal plan for a parent newly diagnosed with diabetes 90.4%. The company writes that its latest models score very highly on many LifeVal tasks.

For a working person in Vietnam, these numbers do not yet answer “can it do my job?” The introduction does not say who or what does the scoring, or what 100% on the home search means: every condition met, or just one house picked. It also says only “our latest models” and names no model. The home-search example is tied to a county in the US, with a transfer-tax calculation and adding the deed to the county’s cart; one of the two underlying surveys is American, and the page does not say which tasks apply to services in Vietnam.

This week’s job is to build a small version of LifeVal yourself. Pick something you repeat, such as comparing three quotes or shortlisting three schools for your child, and write down 4–5 conditions that must hold before you ask the AI. Then check each condition against an official source and record how many held, how many times the AI had to ask you something, and how long it took you to fix the result. That gives you a real number about your own task instead of someone else’s 93.1%.

AWS distills an agent: the small model matches the large one, but no bill yet

The same day, AWS posted a technical walkthrough of knowledge distillation (having a large model do the work, then teaching a small model to imitate its answers), specifically the text-only variant called SeqKD. The example is a shopping agent (AI software that calls tools over many steps). A large model, GLM-5.3, runs on Amazon Bedrock (an AWS service for calling ready-made models) as the agent and generates the working turns; the ShoppingBench (a shopping test set) grader keeps only the turns that solved the task, and those are used to train a small model, Qwen3.5-4B, on Amazon SageMaker AI (an AWS service for training and running models).

On 324 held-out tasks, meaning tasks not used in training, the distilled small model’s average success rate (ASR) is 0.541, against 0.297 before training. The large teacher model reaches 0.511 with random sampling and 0.540 when run in its most deterministic mode; a mid-size frontier model (one of the strongest models available today) reaches 0.546. AWS itself concludes the 4B model recovers most of the teacher’s quality, within run-to-run variation. Keeping the teacher’s reasoning, not just the tool calls, adds roughly 13 points. The price is on IFEval (a test of following checkable instructions, such as “answer in exactly three bullet points”), which drops from 0.819 to 0.741 on the strict prompt-level measure and from 0.872 to 0.813 on the strict instruction-level measure.

The most important part for a decision-maker is at the end of the post. AWS writes that it does not publish cost or latency per task, because that depends on your traffic. A large model called through an API (the door a program uses to call a service) is billed per token (the unit of text a model processes) at every step, while a model you host yourself on a server is billed by the hour whether it serves one task or thousands. So at low traffic, a small model is not necessarily cheaper. AWS also warns that the method works only when the task is within reach of the small model: when the gap is too large, the model copies the form of the answers without solving the task.

If your team pays for an agent running on a large model, do not approve a switch to a small one on a success rate alone. This week’s job: collect 30–50 real tasks with known answers, run both models on them, and record the cost per completed task at your actual traffic. AWS’s 0.541 describes only its own shopping test.

One question for both

The two tests above are this article’s suggestions and have not been run.

Ask whoever hands you a score one question: which part of the measurement is not published? For LifeVal it is the scoring method; for the AWS post it is the bill. The answer often sits inside a small task you can try this afternoon.

Sources

1. OpenAI, “LifeVal” (November 2, 2026) — OpenAI’s own introduction page; this article read only that page.
2. Zappia, Agarwal, Dashti, Oguike, “Build efficient agents with knowledge distillation on Amazon SageMaker AI”, AWS Machine Learning Blog (November 2, 2026) — AWS technical post.