Paper Finds Claude Opus 5.5 Agent Beats Automated CT Labeling Harness

A team of three researchers posted a paper on October 30, 2026, testing whether a large language model labels CT scans of the aorta better as a free-form agent that writes its own code, or inside a harness that automates a first labeling pass and asks the model only to review it. The question matters because dense aortic labels are slow for a human to draw, and the public datasets that could substitute for fresh labeling each use different conventions for what counts as which structure.

The team built a harness called Taurus and tested it against free-form coding agents using blind label recovery: human labels were withheld from six public datasets covering 939 volumes, and every approach was scored against each dataset’s own classes. The first version of Taurus barely helped. The model’s review raised the main metric on only two of the six datasets, by less than the noise between two independent reviews of the same scan. A free-form agent running Claude Opus 5.5 scored higher on most datasets in that first comparison.

The researchers then folded what the free-form agents did well into the automated pass itself, which carried most of the improvement on 38 held-out cases. The finished version smooths label surfaces with a fixed rule and flags the large errors an automated pass tends to make, such as an aneurysm drawn at half its true size. Each flagged case becomes a multiple-choice question put to the model, with candidate corrections already drawn on the scan. The result matched the surface smoothness of human reference labels and showed no difference from the free-form Opus agent on the primary metric, at under half the cost per case.

The numbers

Taurus priced out at $1.22 per case against $2.71 for the free-form agent running Claude Opus 5.5, a gap the authors attribute mostly to the harness and its automation rather than to the model itself. Our price table lists Claude Opus 5 at $5 per million input tokens and $25 per million output tokens, the published rate an agent burns through while it writes and debugs segmentation code case by case, which is the kind of per-token spend that separates a $2.71 agent run from a $1.22 harness pass.

For teams weighing agent based annotation pipelines, the paper’s practical claim is narrower than the headline suggests: most of the gain came from engineering the automated pass, not from swapping in a stronger model. That argues for spending review budget on the harness’s error checks before spending it on a pricier model call, since the model’s own judgment barely moved the outcome once the harness caught the big mistakes on its own.