REFERENCE / CHECKED 16 OCTOBER 2026
Every dollar has
a place in the sum.
Five disjoint token categories. Two rate sets. Three totals that reconcile. This reference makes the comparison reproducible without connecting an account.
01 The rate card
| Category | Fable 5 | Fable 5.1 |
|---|---|---|
| Fresh input | $10 | $10 |
| 5-minute cache writes | $12.50 | $12.50 |
| 1-hour cache writes | $20 | $20 |
| Cache reads | $1 | $0.25 |
| Output | $50 | $50 |
Source: Anthropic's pricing reference. The Fable 5.1 change notes identify the cache-read reduction. Check these sources if you are using this page after the date above.
Scope excludes batch discounts, US-only routing premiums, negotiated rates, partner-cloud billing, tax and server-tool charges. It cannot explain Pro or Max subscription capacity.
02 The counter dictionary
- Fresh input
input_tokensEnter the uncached input counter. Do not enter total prompt tokens here and then add cached tokens again.
- 5-minute writes
cache_creation.ephemeral_5m_input_tokensThe portion created with the shorter lifetime.
- 1-hour writes
cache_creation.ephemeral_1h_input_tokensThe portion created with the longer lifetime. When a source reports only aggregate cache creation, you need the lifetime split before this comparison can price it correctly.
- Cache reads
cache_read_input_tokensKeep reads separate from fresh input and from writes.
- Output
output_tokensUse the reported usage counter. Estimating from visible response text is a different measurement.
The two creation subcounts sum to cache_creation_input_tokens. Do not add that aggregate as a sixth category. Source: Anthropic's prompt-caching usage fields.
Sum counts over the same completed task population for both models. Keep provider, model and pricing mode consistent. A missing field is unknown, not automatically zero. Enter zero only when you know the category was unused.
A missing lifetime is not a zero.
Before entering five counts, reconcile the usage record. This original synthetic aggregate is not a response from an actual model run. The field relationships follow Anthropic's prompt-caching reference.
{
"input_tokens": 1000000,
"cache_creation_input_tokens": 200000,
"cache_creation": {
"ephemeral_5m_input_tokens": 150000,
"ephemeral_1h_input_tokens": 50000
},
"cache_read_input_tokens": 20000000,
"output_tokens": 200000
}The write subcounts reconcile: 150,000 + 50,000 = 200,000. Enter 1,000,000 / 150,000 / 50,000 / 20,000,000 / 200,000 in the five categories, in that order. Do not also add the 200,000 creation aggregate as another category, and do not fold reads into fresh input.
The display rounds the total to $27.88. The same counts at the former cache-read rate total $42.875. The rate-only difference remains exactly $15.
Remove the split and you lose information.
Now imagine the same record includes the aggregate 200,000 writes but omits both lifetime subcounts. At the displayed rates, all-short writes would cost $2.50 and all-long writes would cost $4.00. Holding the other categories fixed gives a token-cost range of $27.50–$29.00. That range expresses the missing lifetime split; it is not a confidence interval or an invoice estimate.
The calculator asks for exact category counts. Obtain the split from the original usage export or explicitly label a scenario assumption. Entering zero for the unknown 1-hour portion silently selects one end of the range.
Stop when the record contradicts itself.
If an aggregate says 200,000 but the two supplied subcounts say 100,000 and 50,000, 50,000 tokens are unaccounted for. Do not allocate them to the cheaper category just to make the form accept the totals. Check whether you mixed records, scopes or incomplete exports. Likewise, an absent output field is unknown, not evidence of zero output.
Download all four synthetic mapping cases · JSON: complete, missing split, inconsistent split and missing output. Expected conclusions are included so the cases can be reviewed without an account or model request.
Provider rates checked 25 October 2026. These calculations cover the token categories shown here; they do not reconcile discounts, tools, taxes, routing premiums or partner invoices. Keep those scopes separate when comparing against an actual bill.
03 The calculation
- A = old counts × old ratesSum across the five categories.
- B = old counts × new ratesA counterfactual: only the price changes.
- C = new counts × new ratesThe second run at the new prices.
Rate effect = B − A. Usage effect = C − B. Together they equal C − A, the net change. A negative effect reduces cost; a positive effect increases it. Percentage change uses A as the denominator and is undefined when A is zero.
Counts are non-negative whole tokens, limited to one trillion per category. Arithmetic uses integer quarter-microdollars; one unit is $0.00000025. Totals remain exact before display. The page rounds money to cents and percentages to one decimal, so displayed subtotals may differ from their rounded total by a cent. JSON downloads retain exact integer totals and eight-decimal dollar strings.
Output headroom divides the rate-only saving by the output rate and rounds down to whole tokens. It holds every original count except output fixed. The tool does not infer cache eligibility, cache lifetime, hit rate, output quality or future token consumption.
Follow the three original worked examples, or change the counts yourself.
04 After a cost comparison
If this comparison informs a model change in your own public application, separately check what a visitor actually experiences. Open the app in a fresh browser context and try the public journey. WebTraceCheck is an optional browser for that step; your own browser can also do it.
If that same app sends a non-sensitive test invitation, CleanTempMail is an optional temporary inbox for inspecting the message. Use a durable address for real accounts and recovery. Neither product calculates AI charges or proves model quality. Both share CacheTilt's operator and neither is required to use this site.
05 Identity, privacy & corrections
CacheTilt is an independent educational calculator, not Anthropic documentation, a billing service or a former organisation's website. Its examples and explanations are original. No model was run to produce the displayed examples.
Entered counts stay in this tab. There is no upload, login, payment, analytics script, cookie or browser storage in the calculator. Reloading restores the synthetic example. Downloads stay on your device. Cloudflare serves the website and processes ordinary request metadata.
Optional product links carry fixed campaign labels for this site and placement. They do not include your counts or results. The destination services have separate policies. For corrections, email support@cleantempmail.com with the rate date and category; do not send credentials or private billing records.