Four models, the same ERR12833637 HiFi library, two execution conditions and three repeats per condition. Scores compare 31-base DNA fragments against the verified reference; they are not binary pass rates.
| Model | Open-ended code mean score | Galaxy-API code mean score | Open-ended code tokens | Galaxy tokens | Galaxy / Open-ended |
|---|---|---|---|---|---|
| gpt-5.5 | 65.95% | 97.10% | 8,932,603 | 6,794,866 | 0.761× |
| gpt-5.6-sol | 98.95% | 100.00% | 25,141,529 | 37,948,455 | 1.509× |
| gpt-5.6-luna | 98.93% | 99.16% | 60,035,833 | 32,720,774 | 0.545× |
| deepseek-v4-pro | 98.98% | 99.62% | 50,968,909 | 6,708,196 | 0.132× |
Tokens are three-run totals. Recompute with bash run.sh, then compare with python3 check.py results reference-results.
This checks saved-answer scoring and token arithmetic. It does not rerun the agents, independently establish the biological reference, or represent an API invoice.