AI Cost Intelligence

Token math, model trade-offs, and the real cost of running AI in production.

Every model provider publishes a per-token rate. Almost nobody publishes what their actual workload costs once you account for retries, context bloat, structured-output schemas, fine-tuning, and the difference between batch and real-time. We do.

Available calculators
3

Tools

Popular model comparisons

Head-to-head API cost on the same real workloads — always current, priced live from our data.

See all 10 comparisons →

How to use this section

Work through the numbers in the order that prevents bad decisions.

  1. 01

    Start with task volume, not model names

    A cheap model can still be expensive if it needs long prompts, retries, or verbose outputs. Use the cost-per-task calculator first to anchor the workload, then compare providers once the input/output shape is realistic.

  2. 02

    Separate prototype cost from production cost

    A prototype usually optimizes for speed and quality. Production needs caching, batching, fallbacks, and a monthly error budget. The API calculator is built to make those operating assumptions visible before the invoice arrives.

  3. 03

    Only rent GPUs after a breakeven check

    Self-hosting looks cheaper per token only when utilization is high enough. If a GPU sits idle, the hourly bill keeps running. Use the self-host vs API page before turning cloud GPU pricing into a commitment.

Guides & deep dives

Long-form explainers that go deeper than the calculators — the hidden traps, the real math, and how operators actually cut the bill.

Why this category exists

  • Pricing changes every quarter

    Anthropic, OpenAI, and Google have each adjusted their pricing structure at least twice in the past 18 months. A calculator that doesn't track these changes is worse than no calculator.

  • Token counts are not intuitive

    A 500-word document is not 500 tokens. Whether you're paying for input, output, or cached context matters enormously. We expose all three.

  • Vendor calculators undercount

    Most vendor estimators show you the per-call cost of a single prompt and stop there. We model your monthly workload including retries, error budget, and the cost of switching providers later.

FAQ

How often is pricing data updated?
LLM pricing is reviewed on a seven-day cadence against public vendor pricing pages. The last completed verification date appears on each tool page.
Do you include cached input discounts?
Yes. Anthropic's prompt caching, OpenAI's batch API, and Gemini's context cache are all modeled in the relevant calculators.
What about open-source models on Together / Replicate?
The calculator includes the open-model providers listed in its current model table. Provider coverage and source links are visible directly in the calculator.

Related