AI Intelligence Calculator

    Methodology

    The calculator compares providers on one objective measure — how many tokens you get per dollar — then weights it by a transparent capability tier. Here is exactly how it works.

    Tokens per dollar

    Every model has two API prices per 1 million tokens: input and output. Because output tokens cost roughly three times as much as input tokens in typical use, we blend them in a 1:3 ratio to get a single effective price per million tokens.

    blended cost = (input × 1 + output × 3) / 4

    tokens per $ = 1,000,000 / blended cost

    The intelligence score

    Not all tokens are equal. We multiply tokens per dollar by a quality factor based on the model's capability tier:

    • Frontier (1.0) — top flagship models (e.g. GPT-5, Claude Opus, Gemini Pro)
    • Balanced (0.8) — strong everyday models (e.g. Claude Sonnet, GPT-4o)
    • Fast (0.5) — small, fast models (e.g. GPT-5 mini, Haiku, Flash)

    The tier is an editorial judgment, not a lab benchmark. It exists so the score is honest about the difference between a fast mini model and a frontier flagship. Each card also shows the raw tokens-per-dollar figure so the two are never conflated.

    The Intelligence Index score

    In addition to our own value score, every model displays an Intelligence Index score from Artificial Analysis— an independent lab that benchmarks models on nine evaluations covering agentic work, coding, tool use, knowledge, reasoning and long-context tasks. The index runs 0–100 and higher is better; for example, Claude Opus 5 scores 63 while Llama 4 Scout scores 10.

    We pull the score published for each model and will link to its Artificial Analysis page so you can inspect the underlying benchmarks. The Intelligence Index measures raw capability and is independent of price — pairing it with tokens-per-dollar shows both sides of the value equation. Not every model on the market is benchmarked, so a small number of rows may not show a score.

    Data freshness

    AI pricing changes frequently. Every model carries a last-verified date and we review the data monthly. Treat all figures as estimates.