RetireBench: how accurate are AI models at retirement math?

RetireBench asks leading AI models the retirement questions people actually face (Social Security claiming, RMDs, Roth conversions, IRMAA surcharges, ACA subsidies, federal tax) and scores the answers against exact, rule-based math.

Preliminary development results. The numbers below come from development runs on RetireBench's public question set (latest run 2026-07-18), rescored under a tolerant answer-extraction policy. They are directionally useful but not official: the audited run on the private held-out set is in preparation, and this page will switch to those results when it lands.

Leaderboard: Core split

Each model answered the same 160 questions: eight domains, twenty questions each, spanning single-rule lookups (T1) through multi-step calculations (T3). Accuracy is equal-weighted across domains. Costs are the metered API cost of the full run.

# Model 95% CI
1 Claude Fable 5 Anthropic 95.6% 91%–98% $4.26 $0.028 10.9 s
2 Claude Opus 4.8 Anthropic 64.4% 57%–71% $1.52 $0.015 5.9 s
3 GPT-5 OpenAI 55.0% 47%–63% $4.53 $0.052 36.6 s
4 Gemini 2.5 Pro Google 47.5% 40%–55% $3.73 $0.049 21.8 s
5 DeepSeek R1 DeepSeek 45.6% 38%–53% $2.62 $0.036 265.6 s
6 GPT-5 Mini OpenAI 43.1% 36%–51% $0.47 $0.0068 19.6 s
7 GPT-5 Nano OpenAI 36.3% 29%–44% $0.27 $0.0046 35.3 s
8 DeepSeek V3.1 DeepSeek 21.3% 16%–28% $0.11 $0.0032 8.6 s
9 Llama 3.3 70B Meta 18.8% 13%–26% $0.020 $0.0007 1.8 s
10 Claude Haiku 4.5 Anthropic 18.1% 13%–25% $0.18 $0.0062 2.5 s
11 Gemini 2.5 Flash Google 16.3% 11%–23% $0.060 $0.0023 1.1 s
12 Qwen3 Max Alibaba 16.3% 11%–23% $0.022 $0.0009 44.0 s
13 Qwen3 235B Alibaba 13.8% 9%–20% $0.0046 $0.0002 1.2 s

= on the accuracy-vs-cost Pareto frontier: no cheaper run scored higher. 95% CI is the Wilson interval on 160 questions. Cost / correct = run cost ÷ correct answers.

Accuracy vs. cost

Getting retirement math right is worth paying for, but the price spread is three orders of magnitude. Gold points are on the Pareto frontier; hollow points are dominated (a cheaper model scored at least as well).

0% 25% 50% 75% 100% $0.01 $0.10 $1 $10 Total run cost, 160 questions (USD, log scale) Accuracy Claude Fable 5: 95.6% accuracy, $4.26 run cost Claude Fable 5 Claude Opus 4.8: 64.4% accuracy, $1.52 run cost Claude Opus 4.8 GPT-5: 55.0% accuracy, $4.53 run cost GPT-5 Gemini 2.5 Pro: 47.5% accuracy, $3.73 run cost Gemini 2.5 Pro DeepSeek R1: 45.6% accuracy, $2.62 run cost DeepSeek R1 GPT-5 Mini: 43.1% accuracy, $0.47 run cost GPT-5 Mini GPT-5 Nano: 36.3% accuracy, $0.27 run cost GPT-5 Nano DeepSeek V3.1: 21.3% accuracy, $0.11 run cost DeepSeek V3.1 Llama 3.3 70B: 18.8% accuracy, $0.020 run cost Llama 3.3 70B Claude Haiku 4.5: 18.1% accuracy, $0.18 run cost Claude Haiku 4.5 Gemini 2.5 Flash: 16.3% accuracy, $0.060 run cost Gemini 2.5 Flash Qwen3 Max: 16.3% accuracy, $0.022 run cost Qwen3 Max Qwen3 235B: 13.8% accuracy, $0.0046 run cost Qwen3 235B
Vertical whiskers are Wilson 95% confidence intervals. Every value is in the leaderboard table.

Where models fail: accuracy by domain

Aggregate accuracy hides the shape of the failures. Models that look passable overall can be coin-flips on IRMAA tiers or ACA subsidy math, the domains where a wrong answer has a real price tag.

Model ACA subsidiesEarly withdrawalFederal taxIRMAARMDsRoth conversionSS claimingTaxable SS
Claude Fable 5 85% 100% 95% 95% 100% 95% 95% 100%
Claude Opus 4.8 75% 85% 70% 5% 65% 20% 100% 95%
GPT-5 15% 100% 25% 0% 100% 0% 100% 100%
Gemini 2.5 Pro 5% 100% 20% 0% 75% 5% 80% 95%
DeepSeek R1 5% 100% 20% 0% 65% 0% 85% 90%
GPT-5 Mini 5% 100% 20% 0% 35% 0% 100% 85%
GPT-5 Nano 10% 100% 25% 5% 30% 0% 70% 50%
DeepSeek V3.1 5% 60% 5% 0% 20% 0% 50% 30%
Llama 3.3 70B 5% 55% 0% 0% 15% 0% 25% 50%
Claude Haiku 4.5 0% 50% 10% 10% 35% 0% 25% 15%
Gemini 2.5 Flash 0% 70% 5% 0% 10% 0% 30% 15%
Qwen3 Max 5% 70% 0% 5% 15% 0% 10% 25%
Qwen3 235B 0% 70% 0% 0% 10% 0% 20% 10%

Twenty questions per domain. Cells below 50% are marked. SS = Social Security.

The Hard split: joint planning problems

Real retirement decisions interact. A Roth conversion changes IRMAA, which changes which claiming age produces the highest lifetime after-tax outcome. The Hard split tests exactly that: T4 chains multiple rules inside one domain; T5 requires jointly optimizing across domains. So far one model has run it.

Claude Fable 5 Accuracy Correct
T4: chained, single-domain 94.4% 68 / 72
Chained conversion 83%
Chained decumulation 100%
Distractor 92%
Multi-year 100%
Optimization 92%
Traps 100%
T5: joint, cross-domain 27.8% 10 / 36
Conversion × claiming 50%
Spend × sequencing 8%
Triple joint 25%

The gap is the finding. Claude Fable 5 scores 95.6% on the Core split and 94.4% on chained single-domain problems, then drops to 27.8% when domains have to be optimized jointly. That is the part of retirement planning where software that prices every interaction on one ledger, rather than reasoning step by step, has a structural advantage.

What RetireBench measures

  • Eight domains, 160 questions (Core). Federal tax, Social Security claiming, taxable Social Security, RMDs, Roth conversion, IRMAA, ACA subsidies, and early-withdrawal rules. Twenty questions each, tiers T1–T3 from single-rule lookups to multi-step calculations.
  • Exact answers, no partial credit. Every question has a rule-derived answer with an explicit numeric tolerance. An answer is correct or it isn't.
  • Refusals and truncations count against accuracy. A model that won't answer a retirement question hasn't answered it.
  • Cost and speed are measured, not estimated. Cost is the metered API cost of the run; latency is the per-question median.
  • Uncertainty is reported. Every headline accuracy carries a Wilson 95% confidence interval. On 160 questions, differences of a few points are noise.

How to read these numbers

  • These are development runs on the public question set, rescored under a tolerant extraction policy that credits correct values in non-conforming formats. The official run uses a private held-out set with full provenance.
  • Models were accessed through a routing provider; the specific serving backend can affect latency and occasionally behavior.
  • One run per model. Models also change under stable names, so treat this as a snapshot dated 2026-07-18, not a permanent verdict on any vendor.
  • Benchmark questions have exact answers by construction. Real retirement situations add ambiguity these scores don't capture. A high score here is necessary for a trustworthy answer, not sufficient.

Why RetireGolden runs this benchmark

RetireGolden's planner doesn't ask a language model to do retirement math. Every number comes from a deterministic, auditable ledger with tested, sourced rules. RetireBench exists to measure the alternative people actually use: asking a chatbot. The results so far say what you'd hope and fear. Frontier models are getting genuinely good at single-rule questions, and still fail the joint optimization problems that decide real money.

Next up: the audited official run, and the flagship comparison, running the same models with and without RetireGolden's planning tools connected, to measure what changes when a model can delegate the math.

RetireBench is a RetireGolden project. Nothing on this page is financial, tax, or legal advice.