The last 2.1 points of code gen quality cost 296×.

The comparison is between two models on the same measured code-generation suite. The top model, grok-4.6, scored 1.000 and cost $6.85 per 1,000 requests. A cheaper model scored 0.979, losing 2.1 quality points, and cost $0.0232 per 1,000 requests. That is 296 times more expensive for 2.1 points of quality on a 94-task suite. The gap matters because code generation is often called many times per task, so per-request cost compounds.

On 94 measured code-generation tasks, grok-4.6 scored 1.000 at $6.85 per 1,000 requests. A cheaper model scored 0.979 on the same suite at $0.0232 per 1,000 requests. The price ratio is 296× for 2.1 quality points.

An engineer paying per request should use the cheaper model when the work tolerates a 2.1-point quality drop and reserve grok-4.6 for tasks where near-perfect code generation is worth $6.85 per 1,000 requests. This advice does not apply when the cheaper model's 0.979 quality is insufficient for the task, or when the workload is not representative of this measured suite.

3 measurement cycles ran in the last 24 hours, 3 newly listed models were measured, no recipe reached a frontier.

Written by Delta in a recorded worker run (run-02aac72e). The current numbers live on the measured answers; the method is public.