Skip to main content
PolicyBench

GPT-6 Astra

OpenAI · AI alone, no tools
Current board, snapshot 2026-09-05
United States rank
#2 of 39
US exact
88.0%
United States

United States benchmark

Exact match
88.0%
Bounded score
96.3%
Parse rate
100.0%
Eligibility flags
962/979

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. Free school meals eligibility100.0%0.7%
  2. Local income tax100.0%0.3%
  3. Person-level CHIP eligibility100.0%0.2%
  4. Person-level Early Head Start eligibility100.0%3.1%
  5. Person-level Head Start eligibility100.0%1.2%
  6. Person-level WIC eligibility100.0%0.3%
  7. Self-employment tax100.0%2.1%
  8. SSI100.0%2.0%
  9. Person-level Medicaid eligibility99.4%30%
  10. Reduced-price school meals eligibility99.0%0.1%
  11. TANF99.0%0.3%
  12. Federal refundable credits98.0%3.7%
  13. Person-level Medicare eligibility91.3%11%
  14. Payroll tax86.0%16%
  15. State refundable credits84.0%0.6%
  16. SNAP83.5%4.1%
  17. Federal tax before refundable credits73.0%19%
  18. State tax before refundable credits65.0%6.0%
Table view (exact, within 1%, outputs), hardest first
ProgramExactWithin 1%n
State tax before refundable credits65.0%73.0%100
Federal tax before refundable credits73.0%77.0%100
SNAP83.5%88.7%97
State refundable credits84.0%84.0%100
Payroll tax86.0%88.0%100
Person-level Medicare eligibility91.3%91.3%172
Federal refundable credits98.0%98.0%100
Reduced-price school meals eligibility99.0%99.0%100
TANF99.0%100.0%100
Person-level Medicaid eligibility99.4%99.4%177
Free school meals eligibility100.0%100.0%100
Local income tax100.0%100.0%100
Person-level CHIP eligibility100.0%100.0%177
Person-level Early Head Start eligibility100.0%100.0%38
Person-level Head Start eligibility100.0%100.0%38
Person-level WIC eligibility100.0%100.0%177
Self-employment tax100.0%100.0%100
SSI100.0%100.0%97

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 8,783 rows whose legacy threshold score is below 1. That universe contains 8,780 of 8,780 exact-match misses and 3 exact hits. Another 1,605 rows with bounded score below 100 were not selected and have no audit annotation.