agents benchmark
Program Bench
Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation. It spans 200 tasks from small CLI tools to large systems such as FFmpeg and SQLite, with submissions judged against more than 248,000 fuzz-generated behavioral tests.
Updated Aug 7, 2026
Published models5
Registry coverage5
MetricScore
EvidenceB
Program Bench leaderboard
Sorted by the source-provided rank. Higher score is better according to the registry.
| 1 | MA | 77.8% | 100.0% | 5 | C | |
| 2 | ZA | 63.7% | 75.0% | 5 | C | |
| 3 | MA | 53.6% | 50.0% | 5 | C | |
| 4 | BY | 50.3% | 25.0% | 5 | C | |
| 5 | BY | 49.4% | 0.0% | 5 | C |
Score distribution
Top published rows on the benchmark's original scale.
Program Bench highlights
The top published results on this benchmark's own scale.
What is Program Bench?
Definition and scoring fields from the benchmark registry.
Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation. It spans 200 tasks from small CLI tools to large systems such as FFmpeg and SQLite, with submissions judged against more than 248,000 fuzz-generated behavioral tests.
Scores are shown in ratio. The current registry marks this benchmark as not independently verified with evidence level B.
- Family
- Program Bench
- Modality
- text
- Primary category
- agents
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- program-bench|llm-stats-current
Source-native results are preserved. Eligibility for the overall LLMBoard score is a separate policy decision.
FAQ
Common questions about Program Bench.
Which model scores highest on Program Bench?
Kimi K3 is currently ranked first with 77.8%.
What does Program Bench measure?
Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation. It spans 200 tasks from small CLI tools to large systems such as FFmpeg and SQLite, with submissions judged against more than 248,000 fuzz-generated behavioral tests.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
5 unique published model results are currently shown.
Does this benchmark affect the overall score?
This benchmark is preserved as source-native evidence but is not eligible for the current overall score.