Skip to content

Value ranking

Best value on MATH-500

As of October 25, 2026, the best value on MATH-500 is Llama 3.3 70B Instruct, at 720.60 normalized points per dollar of input price ($0.10 per million tokens, score 72.1).

500 high-school competition math problems requiring multi-step solutions. Scored on final-answer correctness.

“Value” is normalized benchmark score (0–100 for this leaderboard cohort) divided by input price per million tokens. Higher means more capability per dollar on this axis only — always sanity-check latency, context length, and your real workload.

  1. 1
    Llama 3.3 70B Instruct
    Meta
    720.60
    72.1 / $0.10/M
  2. 2
    Qwen2.5 72B Instruct
    Alibaba (Qwen)
    277.78
    100.0 / $0.36/M
  3. 3
    Llama 3.1 70B Instruct
    Meta
    117.65
    47.1 / $0.40/M
  4. 4
    Mistral Large 2
    Mistral
    37.50
    75.0 / $2.00/M
  5. 5
    Mixtral 8x22B
    Mistral
    0.00
    0.0 / $2.00/M
Open full leaderboard

AI Model Analyzer does not recommend specific vendors; rankings are derived from public data only.