In 10 runs, GPT-6 Luna passed all 178 external acceptance tests in 3 runs without the harness and in all 10 runs with it. The harness supplied API specifications, examples, and deterministic checks that returned feedback during development. Harnessed Luna used more tokens and time, but its average estimated cost was $0.273 per run, versus $0.406 for GPT-6 Sol without a harness—a reported 33% lower cost. GPT-6 Sol passed in all 10 runs in either condition; adding the harness made its runs substantially slower and more expensive without improving the observed pass rate. The experiment concerns one detailed service brief and does not establish how the method generalizes to other tasks.
Cheap, but unreliable. Smaller AI models are tempting for coding tasks because they cost less, and teams rarely choose a model on raw intelligence alone.
In one minute
Naresh Jain reports a controlled comparison of coding agents building the same API service with and without a behavioral harness.
What changed: A coding-agent experiment found that a behavioral harness raised GPT-6 Luna success from 3/10 to 10/10 runs while costing 33% less per run than GPT-6 Sol.