Benchmark Summary

June 17, 2026 ยท View on GitHub

Quick Reference

ModelPass@1Latency (P50)CostDateTier
gpt-5-nano94.5%10.4s~$0.10-0.30/1M2025-10-22Frontier
gemini-3-flash-preview61.6%0.97s$0.00 (free)2025-10-22Mid-Range

Latest Results

gpt-5-nano (2025-10-22) - FRONTIER-TIER PERFORMANCE

Performance:

  • 155/164 tests passed (94.5%)
  • 10.4s median latency
  • ~$0.10-0.30/1M tokens (very affordable)
  • TOP 5 globally

Key Findings:

  • Frontier-tier accuracy competitive with $15-60/1M models
  • Only 9 failures out of 164 tasks
  • Comparable to o1, Claude 3.7 Sonnet, GPT-4.5 Turbo
  • 10-50x cheaper than premium competitors

Full Report: README.md

gemini-3-flash-preview (2025-10-22) - SPEED CHAMPION

Performance:

  • 101/164 tests passed (61.6%)
    
  • 0.97s median latency
  • $0.00 cost (free tier)

Key Findings:

  • All failures were logic errors (no syntax errors)
  • Consistent sub-second response times
  • Suitable for development and prototyping
  • Consider premium models for >70% accuracy needs

Full Report: HUMANEVAL_2025-10-22.md

How to Run

# Full benchmark
make bench-humaneval PROVIDER=gemini MODEL='gemini-3-flash-preview'

# Quick test (10 tasks)
make bench-humaneval PROVIDER=gemini MODEL='gemini-3-flash-preview' N_HE=10

# Generate charts
python3 scripts/generate_benchmark_chart.py reports/HE_*.json

Visualization

ASCII chart available via:

python3 scripts/generate_benchmark_chart.py reports/HE_20251022-135834_gemini-3-flash-preview_tools-0_N164.json

Output:


                        HUMANEVAL BENCHMARK RESULTS                           

  PASS RATE:  61.6%                                                         
                                                                              
  []  
   101 passed     63 failed                                              

Files

  • Detailed Analysis: README.md
  • Methodology: README.md
  • Scripts: ../../scripts/bench_humaneval.py, ../../scripts/generate_benchmark_chart.py