PolicyBench
GPT-5.6 Terra
OpenAI · AI alone, no tools
United States rank
#4 of 23
US exact
83.4%
United States
United States benchmark
Exact match
83.4%
Bounded score
93.7%
Parse rate
100.0%
Eligibility flags
963/984
Score by programhardest first
| Program | Exact | Within 1% | n |
|---|---|---|---|
| State tax before refundable credits | 58.0% | 62.0% | 100 |
| Federal tax before refundable credits | 63.0% | 69.0% | 100 |
| State refundable credits | 78.0% | 78.0% | 100 |
| SNAP | 79.0% | 81.0% | 100 |
| Payroll tax | 83.0% | 85.0% | 100 |
| Federal refundable credits | 90.0% | 91.0% | 100 |
| Person-level Medicaid eligibility | 91.0% | 91.0% | 177 |
| SSI | 96.0% | 97.0% | 100 |
| Free school meals eligibility | 98.0% | 98.0% | 100 |
| Reduced-price school meals eligibility | 99.0% | 99.0% | 100 |
| Self-employment tax | 99.0% | 99.0% | 100 |
| TANF | 99.0% | 99.0% | 100 |
| Person-level Medicare eligibility | 99.4% | 99.4% | 177 |
| Person-level WIC eligibility | 99.4% | 99.4% | 177 |
| Local income tax | 100.0% | 100.0% | 100 |
| Person-level CHIP eligibility | 100.0% | 100.0% | 177 |
| Person-level Early Head Start eligibility | 100.0% | 100.0% | 38 |
| Person-level Head Start eligibility | 100.0% | 100.0% | 38 |
Hardest casesworst misses on positive references
- Federal tax before refundable credits671% offReference $1,910 · predicted $14,728Inspect household #068
- State tax before refundable credits280% offReference $1,253 · predicted $4,764Inspect household #068
- State tax before refundable credits248% offReference $514 · predicted $1,789Inspect household #039
- Federal tax before refundable credits214% offReference $1,346 · predicted $4,227Inspect household #039
- State refundable credits204% offReference $473 · predicted $1,437Inspect household #100
- State tax before refundable credits157% offReference $1,028 · predicted $2,640Inspect household #122
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.