#PolicyEngine
We count amounts within $1 of PolicyEngine's and matching eligibility answers as right, and weight each by household impact. We stop scoring one Minnesota output, whose prompt doesn't say if the renter has a Certificate of Rent Paid. policybench.org/?country=us&scenario=scenario_122#scenarios
PolicyBench, by PolicyEngine
How accurately frontier AI models estimate US tax and benefit amounts, scored against PolicyEngine reference outputs.
policybench.org
November 13, 2026 at 4:29 PM
Our latest post covers the provisions we updated and how the IRS amounts compare with our projections:

www.policyengine.org/us/research/...
Federal tax inflation adjustments for tax year 2027
PolicyEngine has updated its federal tax parameters to the IRS's 2027 inflation-adjusted amounts.
www.policyengine.org
November 11, 2026 at 6:14 PM
Even as Claude Fable 5.2 sets records on major coding and capabilities benchmarks, we find that without tools it still miscalculates taxes, credits and benefits about a third of the time for the households that owe or receive them.
November 13, 2026 at 6:05 PM
Thursday at 10:15am ET at #2026APPAM, Max Ghenis joins Tara Watson (@taraelizwatson.bsky.social) of @brookings.edu, Linda Giannarelli of @urbaninstitute.bsky.social and moderator Cynthia Osborne of Vanderbilt.

Max shows how to compare families' net resources across states with PolicyEngine.
November 5, 2026 at 2:12 AM
We ask AI to compute taxes and benefits for 100 US households and weight each answer by household impact. We count dollar amounts within $1 of what we compute with PolicyEngine, and eligibility answers that match.
November 4, 2026 at 2:04 PM
A video game doesn't guess where a thrown ball lands; its physics engine computes the path. On Tax Chats, I told Jeff Hoopes and Scott Dyreng I expect AI to compute taxes and benefits the same way, by calling an engine that encodes the law: www.buzzsprout.com/1878989/epis...
Max Ghenis: Can AI Correctly Calculate a Tax Liability? - Tax Chats
Jeff and Scott chat with Max Ghenis, co-founder and CEO of PolicyEngine and founder and CEO of the Axiom Foundation, about using technology to improve tax policy analysis. We discuss PolicyEngine, an ...
www.buzzsprout.com
November 2, 2026 at 4:49 PM
GPT-6 Sol leads PolicyBench at 95.2% but gets 67% of nonzero tax and benefit amounts within $1 on its own. In a pilot, we gave it a tool that runs PolicyEngine, the engine behind our reference answers, and it got 97% of those amounts: policybench.org/notes/2026-1...
Given a tool that runs PolicyEngine, three models score above 99%
PolicyBench gave three models a tool that runs PolicyEngine on a household the model writes from the prompt's facts, and ran its 100 US households again on October 14 and 15. GPT-6 Sol, which leads th...
policybench.org
November 2, 2026 at 4:49 PM
Watch our webinar on why estimates for Prop 40, the 2026 Billionaire Tax Act, run from ~$100B in revenue (Galle, Gamage, Saez & Shanske) to a −$39B mean net present value (Jaros & Rauh). We trace about half the gap to future income tax: youtu.be/KLTDkpxKHLs
Why the Prop 40 revenue estimates differ
YouTube video by PolicyEngine
youtu.be
November 2, 2026 at 12:31 PM
David Trimmer covers both estimates, Fernando Hoces de la Guardia breaks down the gap with three open policy analyses, and Max Ghenis runs both sets of assumptions in our open calculator.

Calculator: www.policyengine.org/us/californi...
Reconciliation: fhoces.github.io/opa-prop40/
California Proposition 40 billionaire tax calculator | PolicyEngine
Scores California's Proposition 40 one-time billionaire tax person by person from current Forbes data, with adjustable residency, migration, valuation, timing and income-tax assumptions, including tho...
www.policyengine.org
November 2, 2026 at 5:54 PM
October 31, 2026 at 1:43 PM
Five years ago today PolicyEngine launched in the UK to democratize evidence-based policy. PolicyEngine US followed in 2022. Anyone can change a federal or state tax or benefit rule and see what it means for their household and the country.

Thanks to all who build and use it.
October 16, 2026 at 7:26 PM
Five years ago today we launched PolicyEngine to democratize evidence-based policy. Anyone can change a tax or benefit rule and see what it means for their household and their country.

Thank you to everyone who has built PolicyEngine with us, and to everyone who uses it.
October 16, 2026 at 7:27 PM
We started in the UK, and PolicyEngine US followed the next year. Anyone can read our open-source models, check how they work and build on them.

policyengine.org
October 16, 2026 at 7:27 PM
Five years ago today PolicyEngine launched in the UK to democratize evidence-based policy, so that anyone can change a tax or benefit rule and see what it means for their household and the country.

Thank you to everyone who has built it with us, and to everyone who uses it.
October 16, 2026 at 7:25 PM
GPT-6 Sol still leads at 95.2%. On Oct 13 we also moved our references to policyengine-us 2.41.0, whose fixes let us score again 9 outputs we'd excluded over engine defects. Few models get those right, so every earlier model's within-$1 rate fell, by 0.18 to 0.42 points. policybench.org/notes
PolicyBench notes
Dated records of board changes and findings. Each note names the data release its numbers come from, and a test in the repository checked every number against the frozen snapshot of that release.
policybench.org
October 14, 2026 at 2:58 PM
Our references also moved to a policyengine-us version that fixes the defects behind 14 outputs we’d stopped scoring. Few models get those right, so every earlier score fell, by 0.24 to 1.37 points. policybench.org/notes/2026-10-10-claude-haiku-5-5-joins-the-board
Claude Haiku 5.5 joins the board as the references move to policyengine-us 2.38.6
Claude Haiku 5.5 joined the board on 2026-10-10, bringing it to 47 models. It gets 79.6% of answers within $1, weighted by household impact, #37 of 47 and 4.8 points above Claude Haiku 4.5 (74.8%, #43...
policybench.org
October 10, 2026 at 1:10 PM
SPM thresholds move with consumer spending, not CPI: renters' rose 6.3% in 2025. New from PolicyEngine: a threshold calculator for any metro or state, 2022–2035, every input hashed and reproducible in Python, plus a paper that rebuilds BLS's thresholds from public data to within 3%, most within 2%.
September 15, 2026 at 1:58 PM
Just published in JOSS: 'policyengine: A Microsimulation Tool for Tax-Benefit Policy Analysis' https://doi.org/10.21105/joss.11115
September 14, 2026 at 12:37 PM
Tomorrow in Brussels: the Axiom Foundation and @policyengine.org co-organize a workshop with CAPE and BEAMM — new technologies for evidence-based policy making. Law encoded as executable, cited rules, verified and published in the open.

Free: axiom.org/events/iariw-2026
Workshop: New technologies for evidence-based policy making — Axiom Foundation
Free afternoon workshop in Brussels, Thursday 27 August 2026 — the Axiom Foundation and PolicyEngine with CAPE and BEAMM at UCLouvain Saint-Louis. Talks, a live demo, and a roundtable with Belgian pol...
axiom.org
August 26, 2026 at 6:53 AM
📦 metrictower/funnypot-laravel v0.2.0

Thin Laravel adapter (piece E) over the funnypot-policy decision engine: normalise the request, ask the PolicyEngine, execute the Decision (allow / log / block / deceive). Owns no decision logic.

🔗 https://github.com/metrictower/funnypot-laravel
August 25, 2026 at 11:04 AM
PolicyEngine at @iariw.bsky.social in Brussels this week: two papers, two discussions, and a free workshop Thursday 27 Aug with CAPE and BEAMM — talks, a live demo, a roundtable with Belgian policy institutions.

Register: policyengine.org/iariw-2026
August 23, 2026 at 1:38 PM
New peer-reviewed, open-access paper: PolicyEngine microsimulation helped 39 UK citizens deliberate real fiscal-policy trade-offs during the cost-of-living crisis — a method later adopted by the UK Climate Change Committee. doi.org/10.1016/j.so...
Redirecting
doi.org
June 19, 2026 at 10:47 AM
We just launched a plugin that brings PolicyEngine to Claude.

Describe any tax or benefit reform in plain English. Claude runs a full microsimulation using our open-source model — 700+ variables, 1,200+ parameters, NI to UC.

No code required.

policyengine.org/uk/claude-pl...
February 28, 2026 at 3:59 PM
2025 was a landmark year for PolicyEngine UK.

1,010 local areas modelled
38 reports & trainings published
9 Autumn Budget provisions analysed

From the Local Areas Dashboard to GiveCalc UK—here's what we accomplished.

Explore our interactive year in review: policyengine.org/uk/2025-year-in-review
January 6, 2026 at 5:13 AM