BOZOS / Tools

What does a cheaper
cache read change?

Reprice a fixed workload. Then look at the cost of an attempt and the cost of an observed success.

Presets use the frozen October 9, 2026 audit. Prices are historical assumptions, not a live feed. All calculations happen in your browser.

Your scenario

Sonnet · Medium — original audit inputs.

USD
Total across all attempts
USD per million tokens
USD per million tokens
Include every recorded run
Must be no greater than attempts

The result

Scenario total

$47.64

$21.42 less · 31.01% reduction

Recorded total
$69.05
Cost per attempt
$2.0711
Cost per observed success
$2.8021
Observed successes
17 / 23

Display rounded. All recorded spending, including failed runs, stays in the calculation.

The calculation

Scenario total = recorded total
  − cache-read tokens × (original rate − scenario rate) / 1,000,000

Cost per attempt = scenario total / attempts
Cost per success = scenario total / observed successes

Changing the success count explores a hypothetical denominator. It does not predict whether a rerun will succeed. Editing the recorded cost, tokens, or original rate defines a new baseline; the recorded total must include those original cache-read charges.

Read the result in context

This adjusts one price in the recorded estimate and retains all other costs. It does not reprice cache writes or estimate retry costs, human time, judging, or an actual invoice.

The browser uses floating-point arithmetic and rounds the display. The downloadable audit uses Decimal arithmetic for precise reproduction.

Read the audit and its limits →

Why the denominator matters →