# PlannerBench > Can AI answer retirement questions correctly? PlannerBench is a set of retirement planning questions with known right answers, built by a financial advisor in Sun City West, Arizona. AI models are given the questions and graded on what they say. A benchmark of AI models on 2026 U.S. retirement-planning math (federal tax, Social Security, required minimum distributions, Medicare IRMAA) with an exact answer key, built by Peyton Thalman (https://peytonthalman.com). Every number on the site comes from graded runs and can be rebuilt from the saved replies. Each link below is a Markdown twin of a page, made from the finished page at build time; nothing in it is hidden from the human version. ## Pages - [Home](https://plannerbench.org/index.md): Can AI answer retirement questions correctly? PlannerBench is a set of retirement planning questions with known right answers, built by a financial advisor in Sun City West, Arizona. AI models are given the questions and graded on what they say. - [Results](https://plannerbench.org/results.md): The numbers behind each PlannerBench finding: one rule at a time, several rules at once, the ChatGPT model, the calculator test and the wrong-facts test, with a link to the full detail for each. - [How I built it](https://plannerbench.org/built.md): How a financial advisor, not a programmer, built PlannerBench with AI coding tools: the planning engine, the answer key, a second calculator to check it, secret questions, the calculator test, the wrong-facts test and an outside audit, in plain English. - [Technical details](https://plannerbench.org/method.md): Technical details for PlannerBench: every test and its settings, the GPT-6 Luna run through Codex, how every answer is checked: a second engine written from the rules alone, PolicyEngine-US where it reaches, published tables, a second AI judge and an outside audit, plus the two human checks still owed. - [For AI labs](https://plannerbench.org/labs.md): PlannerBench for AI labs: retirement planning questions with exact, checked answers that a script can grade, where the top models are nearly maxed out, cheap models are not, and none reliably catch wrong facts. - [For financial firms](https://plannerbench.org/firms.md): PlannerBench for financial firms: clients are already asking AI; what the tests found, what fixes it without the most expensive model, and what a firm could do with it. - [AI on its own](https://plannerbench.org/alone.md): How AI models score on retirement questions with an exact answer key: on the one-rule questions tested, Claude Sonnet 5.5 and Haiku 5.5 were almost always right; when several rules meet, scores spread widely; and the 2025 rule many of them misread. - [With a calculator](https://plannerbench.org/calculator.md): Three experiments with a calculator AI models can call while they answer: it fixed one model once it showed the rule, lifted a cheap model above a pricier one, and, for three models, got every whole-household decision right with every number within 2%. It did nothing for others. GPT-6 Luna, the model on the author’s ChatGPT subscription, run through Codex without a calculator, declined those decisions. - [Wrong facts](https://plannerbench.org/wrong-facts.md): I planted one mistake in each of 30 client files, like a retiree who is 14. GPT-6.1 Sol and Claude Sonnet 5.5 caught most of them with almost no false alarms; three of the five cheap and free models caught almost none and gave a confident answer anyway. GPT-6 Luna, the model on my ChatGPT subscription, run through Codex, caught half. - [Example questions](https://plannerbench.org/questions.md): One real PlannerBench question of each of the 14 kinds, with the right answer and how many each AI model got right on questions of that kind. - [Demo walkthrough](https://plannerbench.org/chat.md): A private demo, not open to the public: the planning engine as an add-on for Claude and ChatGPT. It asks one question at a time, does the calculations and says what it assumed. This page shows its real replies to a made-up couple. - [Report](https://plannerbench.org/report.md): The full PlannerBench write-up: AI on retirement math alone, with a calculator, and with wrong facts; method, corrections and limits. ## Data - [models.json](https://plannerbench.org/data/models.json): scores by model with 95% intervals - [tasks.json](https://plannerbench.org/data/tasks.json): scores by task - [mistakes.json](https://plannerbench.org/data/mistakes.json): wrong answers by kind - [tools.json](https://plannerbench.org/data/tools.json): the same hard questions without and with the planning engine's tools - [gauntlet.json](https://plannerbench.org/data/gauntlet.json): whole-household decisions, alone and with the engine - [v2.json](https://plannerbench.org/data/v2.json): bad inputs and the hard tier ## Optional - [The whole site in one file](https://plannerbench.org/llms-full.txt): every page, table and chart's numbers as one Markdown document - [Replays](https://plannerbench.org/replay.md): Watch a model answer a whole-household retirement decision with the calculator: the household, every calculator call, what came back, and the answer against the key. - [Agent mode](https://plannerbench.org/agent.md): PlannerBench as an AI agent reads it: a plain Markdown twin of every page, llms.txt and the whole site as one file, with nothing hidden from the human version.