agents benchmark
Program Bench Leaderboard
Program Bench evaluates code-generation agents on recreating program behavior from a compiled binary and documentation across 200 tasks, with submissions judged against more than 248,000 fuzz-generated behavioral tests.
Updated Oct 23, 2026
Models18
Model coverage18
MetricScore
EvidenceB
Program Bench Ranking
Higher score ranks better on this benchmark.
Sort by
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelAN | Score91.20% | Percentile100.00% | Participants18 | EvidenceC | Evaluated |
| Rank02 | ModelAN | Score82.00% | Percentile94.12% | Participants18 | EvidenceC | Evaluated |
| Rank03 | ModelAN | Score79.70% | Percentile88.24% | Participants18 | EvidenceC | Evaluated |
| Rank04 | ModelMA | Score77.80% | Percentile82.35% | Participants18 | EvidenceC | Evaluated |
| Rank05 | ModelZA | Score63.70% | Percentile76.47% | Participants18 | EvidenceC | Evaluated |
| Rank06 | ModelME | Score59.00% | Percentile70.59% | Participants18 | EvidenceC | Evaluated |
| Rank07 | ModelME | Score59.00% | Percentile64.71% | Participants18 | EvidenceC | Evaluated |
| Rank08 | ModelMA | Score53.60% | Percentile58.82% | Participants18 | EvidenceC | Evaluated |
| Rank09 | ModelBY | Score50.30% | Percentile52.94% | Participants18 | EvidenceC | Evaluated |
| Rank10 | ModelME | Score50.00% | Percentile47.06% | Participants18 | EvidenceC | Evaluated |
| Rank11 | ModelME | Score50.00% | Percentile41.18% | Participants18 | EvidenceC | Evaluated |
| Rank12 | ModelBY | Score49.40% | Percentile35.29% | Participants18 | EvidenceC | Evaluated |
| Rank13 | ModelDE | Score29.10% | Percentile29.41% | Participants18 | EvidenceC | Evaluated |
| Rank14 | ModelXI | Score26.50% | Percentile23.53% | Participants18 | EvidenceC | Evaluated |
| Rank15 | ModelXI | Score26.00% | Percentile17.65% | Participants18 | EvidenceC | Evaluated |
| Rank16 | ModelDE | Score20.30% | Percentile11.76% | Participants18 | EvidenceC | Evaluated |
| Rank17 | ModelZA | Score19.00% | Percentile5.88% | Participants18 | EvidenceC | Evaluated |
| Rank18 | ModelTE | Score17.50% | Percentile0.00% | Participants18 | EvidenceC | Evaluated |
Program Bench Highlights
The leading models and scores on this benchmark.
Program Bench Score Distribution
A closer view of the leading scores on this benchmark.
The Top AI Models for Program Bench
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
What is Program Bench?
What Program Bench measures and how its scores work.
Program Bench is a benchmark for code-generation agents covering tasks ranging from small CLI tools to systems such as FFmpeg and SQLite.
It measures how well agents recreate program behavior from a compiled binary and documentation, using Score reported as a ratio against more than 248,000 fuzz-generated behavioral tests.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- Program Bench
- Modality
- text
- Primary category
- agents
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- overall
LLM Stats. Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about Program Bench.
Which model scores highest on Program Bench?
Claude Opus 5.5 is currently ranked first with 91.20%.
What are the top three models on Program Bench?
The current leaders are Claude Opus 5.5 (91.20%), Claude Haiku 5.5 (82.00%), and Claude Sonnet 5.5 (79.70%).
Which Program Bench model has the lowest official input price?
Claude Haiku 5.5 has the lowest matched official input price at $0.10 input / $0.50 output per 1M tokens.
Which models are fastest among Program Bench results?
The fastest matched records are Claude Sonnet 5.5 (42.00 tok/s via Anthropic), MiMo-V2.6-Pro (13.54 tok/s via Xiaomi), and Claude Opus 5.5 (12.33 tok/s via Anthropic).
Does the highest score result prove overall model quality?
No. This benchmark measures one defined capability or task. The overall LLMBoard score uses a separate aggregation across eligible benchmark evidence.