PolicyEngine US

Mostly multi-step tax math, with one dominant mechanism: 90% of its federal income tax misses compute 2026 under expired TCJA-sunset law — personal exemptions, 10/15/25% brackets, Pease — instead of the extended current law. Gemini shares the issue; GPT-5.5 and Claude mostly don't. policybench.org/?scenario=scen…

@PolicyEngineUS @xai This matters way more than generic benchmarks. Real-world tax computation is harder than a benchmark score. Where's it still struggling?

.@xai's Grok 4.5 joins PolicyBench, our benchmark of how accurately AI computes US taxes and benefits (no tools, scored against PolicyEngine): #5 of 24 — 80.9% of answers within $1 +3.7 points over Grok 4.3 (#9) policybench.org

GPT-5.6 joins PolicyBench, our benchmark of how accurately AI computes US taxes and benefits (no tools, scored against PolicyEngine): Sol #1 of 23 — 88.7% within $1 (prior best 83.5%) Luna #2 — 84.5% Terra #4 — 83.4% policybench.org

See the full benchmark: policybench.org

PolicyBench now scores 20 AI models on how accurately they compute US taxes and benefits — adding Claude Fable 5, Claude Sonnet 5, and five open-weight models: DeepSeek, Qwen, GLM, MiniMax, and Kimi. GPT-5.5 leads, Fable ranks #2, and DeepSeek v4-pro tops the open-weight models.

Federalist No. 62 warned that every new revenue law "presents a new harvest to those who watch the change, and can trace its consequences." 250 years after independence, we build free, open-source software so everyone can trace the consequences. policyengine.org

Thanks to all who joined us. Here's the recording: youtu.be/pSfp7fuMxek

Join us today at 1:30pm ET for a webinar on PolicyBench, our new LLM benchmark that measures accuracy calculating taxes and benefits. us06web.zoom.us/meeting/regist…

We're at the International Microsimulation Association World Congress in Brussels, 1–3 July. Three talks: • Maria Juaristi — local microdata calibration with L0 regularisation • @vahiidahmadi — firm microsimulation + VAT • @MaxGhenis — hands-on PolicyEngine tutorial

A data-quality note. From May 19–June 12, PolicyEngine’s US microdata overstated top-of-distribution aggregates — total AGI, high incomes, itemized/charitable deductions — due to a calibration change. Society-wide results were affected; single-household calculations weren’t. If you ran society-wide analysis between May 19 and June 12 — especially anything sensitive to high incomes or deductions — we recommend re-running on current data. For more on what happened, which surfaces were affected, how to inspect our calibration, and what we're doing to ensure the highest quality data, read our postmortem: github.com/PolicyEngine/p… We're sorry to anyone whose analysis this affected. DMs open if we can support any updating of your work.

You can view the impact of this enacted CTC on our RI CTC Calculator and see how it would affect your household. policyengine.org/us/rhode-islan…

In 2027, we project the CTC will: - Cost $45 million - Benefit over 79,000 households - Lower child poverty and deep child poverty by 3.5% and 8.4%, respectively

Last week, @GovDanMcKee signed the Rhode Island Fiscal Year 2027 budget, which includes a new refundable Child Tax Credit. Beginning in 2027, taxpayers can claim $330 for each child under the age of 19. Using our RI CTC calculator, we analyzed the impact of this program.

The @BEA_News — the federal agency behind GDP — used PolicyEngine to calculate the taxes in its latest Distribution of Personal Income statistics. The first BEA release built on our open-source model. bea.gov/data/special-t…

Want to see it live? We’re walking through PolicyBench on Thursday, June 25 at 1:30pm ET — demo + Q&A. Register → us06web.zoom.us/meeting/regist…

Everything is open: a live leaderboard, a scenario explorer with every prompt and reference output, and a full paper. policybench.org Paper: policybench.org/paper

Can a language model compute a household’s taxes and benefits from the prompt alone — no tools, no lookups? We tested 13 frontier models against PolicyEngine on 100 representative US households and 18 tax and benefit outputs. New benchmark: PolicyBench. policybench.org

Earn a dollar more, take home less. Benefit cliffs happen when a household loses more in benefits than it gains in earnings. CliffWatch shows where, when, and how — across earnings levels and US states. Live demo Fri May 29, 1pm ET → us06web.zoom.us/meeting/regist…