gpt 5.6 "knowledge per $"

Introduction.

I got tired of opening benchmark after benchmark just to answer one question: which model gives me the best quality for my budget? This page compresses hundreds of benchmark points into a decision table you can scan in seconds.

Every benchmark release comes with a wall of bar charts, a dozen sub-scores, and zero guidance on what any of it means for your actual use case. I don't need to know a model's MMLU score — I need to know: for what I'm doing, what's the cheapest model that won't embarrass me? So I turned the leaderboard into something scannable: pick your task complexity on the right, and see the cheapest tier that clears the bar.

Table:

Rows are task complexity, columns are pricing tiers. Find your task, read across, stop at the first green cell — that's the cheapest model that'll do the job well.

The real story here isn't Sol vs Terra vs Luna — it's that Luna is criminally underpriced. At $1/$6, it's a fifth of Sol's cost, and it's fine for most real work. You're paying 5x more on Sol for headroom you'll use maybe 10% of the time. Unless you're doing genuinely hard multi-step reasoning, Luna is the model. Everything above it is you paying for a ceiling you're not touching.

Source: Artificial analysis

Want a full version of this?

This table took me an afternoon and covers one model family. If there's interest, I'd like to build this into a proper site — same idea, dead simple, updated as new models ship, covering every major model on "knowledge per $" instead of raw benchmark scores. If that's useful to you, share this post and drop a comment saying so — that's genuinely what decides whether I build it next.

until them keep vibing . it's founder mode.