First implication: A team evaluating smaller coding models should compare repeated task completion with and without task-specific feedback, not just model price or raw token use.
Thesis
A behavioral harness can make a cost-optimized coding model more dependable on a bounded task, but its value depends on the model and the task’s feedback surface.
Observed facts
GPT-6 Luna passed all 178 external acceptance tests in 3 of 10 unharnessed runs and 10 of 10 harnessed runs.
1GPT-6 Sol passed in all 10 runs with and without the harness; the harnessed runs were substantially slower and more expensive.
1Harnessed GPT-6 Luna averaged an estimated $0.273 per run versus $0.406 for GPT-6 Sol without a harness, a reported 33% lower cost.
1
Mechanism
Machine-readable API specifications and examples guided development, while deterministic checks and test feedback acted as sensors during coding.
Why now
Teams are weighing less costly coding models, and this experiment measures repeated-run reliability and cost rather than relying on a single demo.
Implications
A team evaluating smaller coding models should compare repeated task completion with and without task-specific feedback, not just model price or raw token use.
Evidence boundary
This is one experiment on one unusually detailed service brief; it does not show that harnesses will improve every model, task, or production outcome.
Watch conditions
Whether similar repeated-run comparisons on less complete briefs and different software tasks preserve the reliability and cost advantage.