Skip to main content
PolicyBench

GPT-5.4 mini

OpenAI · AI alone, no tools
Current board, snapshot 2026-11-11
United States rank
#49 of 52
US exact
73.3%
United States

United States benchmark

Exact match
73.3%
Bounded score
80.6%
Parse rate
100.0%
Eligibility flags
891/975

Exact match by programall 18 output groups, best first

Bar length is the unweighted exact-match rate on that program's outputs; the right column is the program's share of the headline weight, which follows household dollars, so the bars do not average to the headline.

  1. Local income tax100.0%0.3%
  2. Reduced-price school meals eligibility100.0%0.1%
  3. TANF99.0%0.3%
  4. SSI97.9%2.0%
  5. Person-level Medicare eligibility97.1%11%
  6. Free school meals eligibility95.0%0.7%
  7. Person-level Early Head Start eligibility92.1%3.1%
  8. Person-level CHIP eligibility92.1%0.2%
  9. Self-employment tax90.0%2.1%
  10. Person-level WIC eligibility86.4%0.3%
  11. Person-level Medicaid eligibility84.4%30%
  12. Person-level Head Start eligibility84.2%1.2%
  13. State refundable credits81.6%0.6%
  14. Federal refundable credits81.0%3.7%
  15. SNAP79.6%4.1%
  16. State tax before refundable credits60.2%6.0%
  17. Payroll tax56.3%16%
  18. Federal tax before refundable credits54.8%19%
Table view (exact, within 1%, outputs), hardest first
ProgramExactWithin 1%n
Federal tax before refundable credits54.8%54.8%84
Payroll tax56.3%67.7%96
State tax before refundable credits60.2%60.2%93
SNAP79.6%79.6%93
Federal refundable credits81.0%81.0%100
State refundable credits81.6%81.6%98
Person-level Head Start eligibility84.2%84.2%38
Person-level Medicaid eligibility84.4%84.4%173
Person-level WIC eligibility86.4%86.4%177
Self-employment tax90.0%95.0%100
Person-level CHIP eligibility92.1%92.1%177
Person-level Early Head Start eligibility92.1%92.1%38
Free school meals eligibility95.0%95.0%100
Person-level Medicare eligibility97.1%97.1%172
SSI97.9%97.9%97
TANF99.0%99.0%100
Local income tax100.0%100.0%98
Reduced-price school meals eligibility100.0%100.0%100

Hardest casesworst misses on positive references

Scores are from the frozen manuscript snapshot under the AI-alone condition: structured responses over the same household facts (whole-scenario or in output subsets, per the serving-configuration table), no tools, graded against PolicyEngine reference outputs. See the leaderboard and paper for methodology. The frozen US annotations cover 9,132 rows whose legacy threshold score is below 1. That universe contains 9,128 of 9,128 exact-match misses and 4 exact hits. Another 2,497 rows with bounded score below 100 were not selected and have no audit annotation.