Cost and Token ManagementPricing and the multiplier stack
No narration yet
Module 4, Lesson 124 min

Pricing and the multiplier stack

Cost and Token Management is only 2.8 percent of the exam, which is about one or two questions. They will be arithmetic questions with exact numbers in them, so they are free marks if you learned the table and lost marks if you did not. Twenty minutes now, permanently.

The per-million-token table

Input first, output second, dollars per million tokens.

Fable 5 and Mythos 5: 10 and 50. The premium tier. Mythos is invite-only.

Opus 5, and the whole Opus 4.x line: 5 and 25. Opus 5 costs the same as its predecessors, which is a nice detail: upgrading within Opus is free at the price level.

Sonnet 5: 2 and 10 through 2026-08-31, then 3 and 15.

Sonnet 4.6 and Sonnet 4.5: 3 and 15.

Haiku 4.5: 1 and 5.

Notice the pattern under all of it: output is five times input, everywhere. Fable 10 to 50, Opus 5 to 25, Sonnet 3 to 15, Haiku 1 to 5. That regularity is easier to remember than ten separate numbers, and it tells you that verbosity is a cost problem five times bigger than context is.

The multipliers

Now the layer on top. These are the numbers that turn a price into a bill.

Batches: minus 50 percent. The single biggest lever available, and the only cost of it is that the work stops being interactive.

Cache write: 1.25 times for a 5-minute TTL, 2 times for a 1-hour TTL. You pay a premium to put something in the cache.

Cache read: 0.1 times. You pay a tenth to read it back. This is where the money is.

Long context: no premium. Deliberately in the list as a non-multiplier, because people assume one exists.

inference_geo: "us": 1.1 times. Data residency costs 10 percent.

Managed Agents: plus 0.08 dollars per session-hour, on top of tokens. Note the unit. This is not a token multiplier, it is a flat rate against wall-clock session time, so a slow session costs more than a fast one even at identical token counts.

Web search: 10 dollars per 1000 searches. Web fetch: free.

Token multipliers versus flat charges

The whole trick to these questions is sorting the list into two piles before you touch any arithmetic. Some of these numbers scale with how many tokens you moved. Others scale with something else entirely, and multiplying them by your token cost gives you a confidently wrong number.

The practical stacking order for a big offline job: pick the cheapest model that meets the requirement, cache the stable prefix, batch the whole thing, and turn effort down if the task is simple. Four levers, all independent, and a workload that uses all four is not in the same universe as one that uses none.

Practice

Try it yourself

Recall

The price table, cold

Straight recall. These numbers are the most lookup-able thing on the exam, which is exactly why getting them wrong is unforgivable.

State the per-million-token input and output prices for Fable, Mythos, Opus 5, Opus 4.x, Sonnet 5, Sonnet 4.6 and 4.5, and Haiku 4.5.

Quiz

The Batches discount

Which multiplier does the Batches API apply?

Quiz

The data residency multiplier

You set inference_geo to us for a compliance requirement. What does it do to your bill?

Quiz

Add up one month of a Managed Agents workload

A month of usage, on claude-sonnet-5 at its current promotional rate, before 2026-08-31.

input tokens      2,000,000
output tokens       500,000
session-hours            20
no caching, no batching, no data residency

Sonnet 5 is 2 dollars per million in and 10 dollars per million out. Managed Agents adds 0.08 dollars per session-hour.

Quiz

Confirming a rate that has a date on it

You are sizing a budget that runs into September and you need to know what Sonnet 5 costs after the promotional period ends. You have the documentation and the console open. Where do you get the answer?

Check

Build a one-line cost model

Estimate the monthly cost of one workload you can describe.

You should see

A single expression of the form (input tokens times input price plus output tokens times output price) times volume, with every applicable multiplier named and applied, and you can state which multipliers are token multipliers and which are flat per-unit charges.