Question 1

Which LLM API is cheapest right now?

Accepted Answer

Sort the table by estimated monthly cost after setting your workload. The lowest estimated cost card surfaces the lowest-cost variant whose quality percentile is at least 40, so untested models cannot win on price by being unmeasured.

Question 2

How accurate is the estimated monthly cost?

Accepted Answer

The token-rate estimate uses input tokens times input rate plus output tokens times output rate, with cached input billed at the provider's cache rate when published. It does not include rate-limit overage, fine-tuning surcharges, or batch discounts.

Question 3

Why does quality only show a fraction like 8/14?

Accepted Answer

The fraction shows benchmark coverage. A score backed by 14 of 14 benchmarks is more reliable than one backed by 3 of 14. Variants with sparse coverage are real estimates, but should be read with lower confidence.

Question 4

Are open-source models actually cheaper?

Accepted Answer

Hosted open-weight models on providers such as Together, Fireworks, or Groq are often much cheaper per token than frontier proprietary APIs at similar quality bands. Self-hosting changes the math because GPU rental and operations time replace API margin.

Question 5

How is the value score calculated?

Accepted Answer

Value is a general index blending normalized quality and cost signals. It is not literal quality per dollar and does not respond to your token workload. It is a shortlisting heuristic, not a substitute for testing on your own prompts.