A general coding agent finished all 244 bioinformatics tasks and scored 0.869, beating specialist agents at 0.758

A general frontier model in an off-the-shelf coding harness can complete every task in a bioinformatics benchmark that requires free-form output files, and do so more accurately than three purpose-built agents. The original PromptBio-Bench evaluation ran Biomni, STELLA and ToolsGenie on Claude Sonnet 4.6, found accuracies of 0.72 to 0.76 that declined with difficulty, and noted that native coding harnesses such as Claude Code or Codex were not evaluated. Under the original framework, Claude Opus 5.5 with skills had completion 1.00 and accuracy 0.869, against 0.758 for Biomni and ToolsGenie and 0.717 for STELLA; paired McNemar tests gave 39 wins to 12 losses against Biomni (p = 2.0 × 10⁻⁴), 38 to 11 against ToolsGenie (p = 1.4 × 10⁻⁴) and 50 to 13 against STELLA (p = 3.0 × 10⁻⁶).

The advantage concentrates on medium- and high-difficulty tasks and on the bioinformatics track, while on the data-science track the specialists were already close to saturation. All three specialists plateaued at 0.72 to 0.73 on medium tasks and 0.56 to 0.69 on high-difficulty tasks. Claude with skills reached 0.878 on medium and 0.756 on high difficulty; track A accuracy was 0.817 against 0.618 to 0.641 for the specialists, and on track B, where Biomni already reached 0.920, Claude reached 0.929; a per-task oracle taking the best specialist result, which no real system achieves, reaches 0.889, two points above Claude alone.

The skill catalogue added little because the agent never loaded a skill. Any difference between the skills and no-skills configurations could have been attributed to the content of 667 bioinformatics skills. The skill catalogue was present in the context of 441 of 444 subagent sessions, yet no session invoked a skill or read a skill file; under the fixed evaluator, the skills configuration reached 0.918 with the GPT judge and 0.926 with the Claude judge, the no-skills configuration 0.873 and 0.902, a gain of 4.5 points with the GPT judge (95% CI +2.0 to +7.4) and 2.5 points with the Claude judge (+0.4 to +4.9); the skills run consumed 101.0 M tokens against 74.6 M without skills.

Scorer and reference defects in the original framework explain more of the failures than the agent's own analytical errors. Under the original framework, 32 tasks scored below 0.5 for Claude with skills, and only 8 of them were the agent's own mistakes or poor choices. Of those 32, 16 were scorer defects, 3 were reference defects and 5 depended on settings the question never states; after eleven evaluator defects were fixed, 18 tasks still failed under the Claude judge, of which 7 are reference defects, 4 unstated settings and 7 agent choices.

The result covers the 244 tasks of PromptBio-Bench, one attempt per task with a 3,600-second limit, with agents free to install tools in a shared conda environment and consult the web; specialist scores come from the original release, so only the original-framework comparison is like for like.

The specialist agents were not re-scored with the fixed evaluator and their outputs were unavailable; the general and specialist systems used different model generations; each configuration was run once, so run-to-run variance is unquantified; the two runs received per-task prompts written by their own orchestrating sessions, so the skills comparison is not controlled at the prompt level; the GPT judge mixed GPT-5.6-sol and GPT-6.1-sol, and the last fixes were not applied to the GPT-judge run; the evaluator fixes and two scoring rules were designed while inspecting one agent's outputs and were not pre-registered; contamination of the public benchmark materials in training data cannot be excluded.