That presumes that there is a linear scale that measures performance. This can be tested: https://en.wikipedia.org/wiki/Rasch_model https://en.wikipedia.org/wik

That presumes that there is a linear scale that measures performance. This can be tested: https://en.wikipedia.org/wiki/Rasch_model https://en.wikipedia.org/wiki/Rasch_model Even assuming this holds, what utility you gain by the best models depend completely by your workload. If you have tasks that require performance 10 and DeepSeek has 9, you will gladly pay for SotA models.