GPT-6 Luna $20 vs Fable 5.1 $2,000 in monthly token-cost test

A short Python script turns published per-token prices into a monthly bill, and the ordering it produces is not intuitive from any single vendor's price list. GPT-6 Luna, at $20 a month, sits an order of magnitude below the next tier: Gemini 3.8 Flash at its introductory price is $150, GPT-6 Sol and Claude Sonnet 5.5 both land at $400, and Claude Opus 5.5 at $800. These figures only count input and output tokens at list rates — no cache discounts are applied, and the same input volume is assumed for every model. That limitation is explicit: the script shows unit price, not the cost of a single task, and it recommends replacing the assumed token volumes with real numbers measured from a user's own logs.

That distinction matters because the script and CursorBench 4.0 tell different stories. CursorBench reports both a score and an average cost per task, and Gemini 3.8 Flash consumes an average of 162,565 tokens per task — enough that, despite a unit price below half of Claude Sonnet 5.5's, its cost per task is higher on that benchmark. The same benchmark shows Claude Opus 5.5 scoring 1.8 points apart between its Max and High configurations while costing roughly 3.4 times as much. Google itself notes that hard tasks can use more tokens and tool calls, and that thinking tokens are billed as output.

The recommended savings sequence follows from that: pick the model tier first, then reduce the tokens being sent, then apply caching. The last step has the most concrete numbers attached — published information says cache reads for GPT-6 Sol and Luna are discounted by 90%, and Claude Opus 5.5 and Sonnet 5.5 charge 5% of the input rate on cache hits. Those discounts only apply to work that resends the same preamble, such as a system prompt or shared document, which is why caching is framed as something that changes the unit price rather than the volume.