THE AI ECONOMICS LAB

A million calls.
What’s the real cost?

An inexpensive call can become an expensive workflow. Explore what changes when you account for every step, every retry, and the outcomes that actually work.

START WITH A WORKLOAD

Illustrative scenarios.
Replace with your measured data.

Your workload

Average usage across all tasks, including unsuccessful ones.

Your model prices

USD per million tokens. Example rates, not vendor quotes.

InputOutputPrimaryAlternative
Explore a different setup

Share of input tokens billed at the cached rate.

Replaces primary calls; assumes equal token usage.

Does it still work?

End-to-end success rates after retries. Supply these from your evals; they are not predicted by this tool.

YOUR PROPOSED SETUP · USD / MONTH

—

Compared with — for your baseline

Calculating…

Cost / successful task—Proposed
Cost / successful task—Baseline

Where the spend goes

Baseline
Proposed
Input tokensOutput tokens
Monthly model calls
—
Successful tasks · proposed
—

Model token costs only. Tools, infrastructure, human review, taxes, and cache-write charges are excluded.

Calculations stay in your browser. Sharing puts these numeric assumptions in the link. No account or API key needed.

READ THE NUMBERS WELL

Cheap tokens aren’t
the whole story.

Steps multiply. So do retries.

A task with four model calls and 20% extra retry calls uses 4.8 calls on average. Multiply the whole workflow by your volume before comparing prices.

Success belongs in the denominator.

Lower spend is useful only in context. Compare monthly cost divided by successful tasks, then consider the cost of failures and human intervention separately.

Caching and routing are hypotheses.

A smaller model may need more tokens or retries. Cache eligibility varies. Validate quality and usage on real traffic before treating an estimate as a saving.

See the assumptions & calculation method

Baseline: all calls use the primary model at the normal input and output rates, with no caching. The proposed setup uses your routing and caching assumptions. Both process the same workload and retry volume.

Monthly calls = tasks × calls per task × (1 + extra retry calls ÷ 100).

Blended price = primary price × (1 − routed share) + alternative price × routed share, calculated separately for input and output.

Proposed input cost = calls × input tokens ÷ 1,000,000 × blended input price × (1 − cached share × cache discount).

Proposed output cost = calls × output tokens ÷ 1,000,000 × blended output price. Output is never discounted by input caching.

Cost per successful task = monthly model cost ÷ (tasks × success rate). If no tasks succeed, this metric is undefined.

Rates are illustrative, in USD, and fully editable. Cached share means tokens actually billed at a cached rate, not cache-eligible requests. The same discount applies to both models. Routing replaces calls without extra classification or fallback calls. There are no tiered, batch, long-context, or volume discounts. Costs scale linearly; this is a planning model, not a billing forecast.

Proposed / month —View results