basforge

Search titles, text and the “If it doesn’t work” cases.

    RU

    Price a model by the solved task, not the token

    A cheap model that needs three attempts and your correction costs more than a strong one that got it right the first time. Compare models and effort levels on your own tasks by cost and time to a finished result — and often a strong model on low effort beats a weak one on high.

    What it looks like

    Finding out why a build fails — 30,000 input tokens (code and log):

    Option Attempts to a correct answer Tokens in total Cost Your time
    Haiku 5.5 4 plus a hint ~140K in, 8K out ~$0.02 20 minutes checking wrong answers
    Sonnet 5.5, high 1–2 ~60K in, 10K out ~$0.22 5 minutes
    Opus 5.5, low 1 ~30K in, 3K out ~$0.18 2 minutes

    The attempt counts are an example, not a measurement: every task has its own. The point is that tokens here cost cents, while checking a wrong answer costs minutes of your work.

    How to compare on your own tasks

    1. Take 5–10 typical tasks with a known correct result.
    2. Run them on two or three options: the small model, the mid-tier, the flagship on low.
    3. Use usage (Tokens) to count the cost up to a correct answer, and note where you had to ask again.
    4. For each kind of task, keep the cheapest option that solves it on the first try.

    Where cheap does pay off

    Thousands of uniform requests with a checkable result: classification, field extraction, translating short strings. There a mistake is caught automatically, and a 20–40× price difference multiplies by the volume. That’s also where batching (Non-urgent API requests go as a batch at half price) and caching (Prompt cache) come in.