The questions are different
Artificial Analysis publishes a composite intelligence score based on its selected evaluations. Arena rankings come from people comparing anonymized model responses. The first emphasizes performance on a benchmark set; the second measures preference among the responses and voters represented in Arena. Their scales are not interchangeable.
What a high score can hide
A model may excel at reasoning benchmarks and still perform poorly on your support queue, extraction schema or style requirements. Arena votes may favor fluency or formatting that your application does not value. Scores also change with dataset updates, model versions and ranking methodology. Read each project's methodology and check the date of any chart.
Use rankings as a shortlist
Select a few models that meet your requirements for latency, context, region and supported modalities. Run the same representative tasks on all candidates. Score pass/fail outcomes with a rubric; where judgment matters, blind the evaluator to model identity. Track token usage and retries so a public ranking does not obscure your operating cost.
Price the winner using your workload
Input and output volumes may differ substantially between models, and a model that needs more retries can cost more than its published token rate implies. Compare model prices, then compute cost per successful task from your own evaluation. Treat benchmark rank as one piece of evidence, not a purchasing decision.