Picking an AI model for bioinformatics usually starts with accuracy. The ten best configurations on VariantBench, a benchmark of genetic variant interpretation, all pass between 40% and 50% of attempts. That is a narrow band. The cost of getting there is not narrow at all.

We took the published cost per attempt for those ten configurations and divided it by their pass rate. The result is the expected spend to obtain one correct answer, assuming attempts are independent.

What does a correct answer cost?

Source: VariantBench leaderboard (LatchBio), sampled October 3, 2026, via benchmarklist.com. Top ten by pass rate out of 33 configurations. The leaderboard does not publish confidence intervals.

View data

config	cost
GPT-6 Astra / Pi (47.7%)	0.67
Grok 4.6 / Grok Build (46.0%)	2.13
Claude Opus 5 / Pi (49.4%)	2.18
GPT-6 Astra / Codex (48.6%)	3.56
Claude Opus 5 / Claude Code (49.7%)	5.59
GPT-5.6 Sol / Codex (42.1%)	5.65
GPT-5.6 Sol / Pi (41.2%)	6.57
Claude Opus 4.8 / mini-SWE (40.7%)	6.69
Claude Opus 4.8 / Pi (42.1%)	7.41
Claude Opus 4.8 / Claude Code (39.8%)	8.49

GPT-6 Astra running in the Pi harness passes 47.7% of attempts at $0.32 each, which works out to $0.67 per correct answer. Claude Opus 5 in Claude Code passes 49.7% at $2.78 each, or $5.59 per correct answer. The accuracy gap between those two is two points. The cost gap is more than eightfold.

Two points of accuracy are within the noise on a benchmark of 118 evaluations. LatchBio's VariantBench paper reports 95% confidence intervals of about eight points around a 42% pass rate. The cost gap, in contrast, is far outside any plausible noise.

Is it the model or the harness?

The same model costs different amounts in different harnesses, which are the agent frameworks wrapped around it. We compared six models that appear in both the Pi harness and their developer's own.

Pi harnessDeveloper's own harness

Source: VariantBench leaderboard (LatchBio), sampled October 3, 2026, via benchmarklist.com. Pass rates differ slightly between harnesses; Grok 4.6 passes 38.1% in Pi and 46.0% in Grok Build.

View data

model	pi	own
GPT-6 Astra	0.32	1.73
Claude Opus 5	1.08	2.78
Grok 4.6	0.49	0.98
Claude Sonnet 5	1.95	2.97
Claude Opus 4.8	3.12	3.38
GPT-5.6 Sol	2.71	2.38

For GPT-6 Astra, the cost per attempt is $0.32 in Pi and $1.73 in Codex, while pass rates are 47.7% and 48.6%. For Claude Opus 5 it is $1.08 against $2.78, at 49.4% and 49.7%. In those cases the more expensive harness bought less than one point of accuracy.

Grok 4.6 is the counterexample. Its own harness costs twice as much ($0.98 against $0.49) and passes 46.0% of attempts against 38.1% in Pi, an eight-point gain. GPT-5.6 Sol is cheaper in its own harness. The harness is a cost lever, and what it buys varies by model.

What this does not show

The leaderboard was sampled on one day and contains no confidence intervals. Cost per attempt reflects each lab's pricing and token use at that time, and prices change. Our cost per correct answer assumes independent attempts, which may not hold when a user retries a hard case. Claude Opus 5.5 does not appear in this leaderboard, so it is not part of this comparison. Our benchmarks dashboard shows pass rates across all three LatchBio benchmarks, but only VariantBench has cost data.

Sources

1. LatchBio, VariantBench, and the leaderboard on benchmarklist.com: pass rates, average cost per run, sampled October 3, 2026 (33 configurations, 118 evaluations, 354 attempts each).
2. Cost per passing attempt and the harness comparison are calculated by the editors from the leaderboard figures.