PolicyBench
GPT-5.6 Luna
OpenAI · AI alone, no tools
United States rank
#2 of 23
US exact
84.5%
United States
United States benchmark
Exact match
84.5%
Bounded score
92.5%
Parse rate
100.0%
Eligibility flags
957/984
Score by programhardest first
| Program | Exact | Within 1% | n |
|---|---|---|---|
| State tax before refundable credits | 53.0% | 56.0% | 100 |
| Federal tax before refundable credits | 65.0% | 70.0% | 100 |
| SNAP | 78.0% | 81.0% | 100 |
| State refundable credits | 79.0% | 79.0% | 100 |
| Payroll tax | 89.0% | 91.0% | 100 |
| Federal refundable credits | 90.0% | 90.0% | 100 |
| Person-level Medicaid eligibility | 93.8% | 93.8% | 177 |
| Person-level CHIP eligibility | 96.6% | 96.6% | 177 |
| Reduced-price school meals eligibility | 97.0% | 97.0% | 100 |
| SSI | 97.0% | 98.0% | 100 |
| Person-level Medicare eligibility | 97.2% | 97.2% | 177 |
| Free school meals eligibility | 98.0% | 98.0% | 100 |
| Self-employment tax | 98.0% | 99.0% | 100 |
| TANF | 99.0% | 99.0% | 100 |
| Local income tax | 100.0% | 100.0% | 100 |
| Person-level Early Head Start eligibility | 100.0% | 100.0% | 38 |
| Person-level Head Start eligibility | 100.0% | 100.0% | 38 |
| Person-level WIC eligibility | 100.0% | 100.0% | 177 |
Hardest casesworst misses on positive references
- SNAP675% offReference $461 · predicted $3,576Inspect household #023
- Federal tax before refundable credits611% offReference $1,910 · predicted $13,573Inspect household #068
- SNAP330% offReference $288 · predicted $1,236Inspect household #108
- State tax before refundable credits257% offReference $1,253 · predicted $4,479Inspect household #068
- State tax before refundable credits242% offReference $514 · predicted $1,760Inspect household #039
- Federal tax before refundable credits233% offReference $1,346 · predicted $4,485Inspect household #039
Scores are from the frozen manuscript snapshot under the AI-alone condition: one structured response per household, no tools, graded against PolicyEngine reference outputs. See the leaderboard and paper for methodology.