Context Carry-Cost Calculator

A block of tokens that enters a session gets re-sent on every inference step after it, every tool call included, not just the messages you type. It costs a fixed amount to bring in, then keeps costing turn after turn until the session ends.

The same read, entered early instead of late in a long session, can cost an order of magnitude more for the identical block. Set the shape of a session below and see what one read actually costs, and how much of that is just about when it arrived.

Session shape

Standing context

Already in the session before the read arrives (rules, prior turns).

Size of the read

Tokens the block adds when it enters.

Turn it enters

Which inference step this block first shows up on.

Total steps in the session

Every tool call is an inference step, not just user messages.

Cache pricing

Cache read multiplier

Price of a cached read, as a fraction of the base input price.

Cache write multiplier

2× on 1-hour TTL.

Model & price

Model preset

Price per MTok, input

Price per MTok, output

Not used below. Carried context is always billed at the input rate.

Carried tokens

—

Cost of this block

—

write once, re-read every later turn

Same block at turn — instead

—

everything else held equal

Position multiplier

—

Cost by entry turn

The same block, priced as if it had entered at every turn from 1 to 300.

entered at the turn set above entered 20 turns from the end instead

Measured on my own workspace, Aug 2026

What this looks like at scale

A 104k-token read at turn 3 of a long session carried 19.6M tokens of re-reading. The same-size read at turn 182 carried 133k, a 147× difference from position alone. Across that session, output tokens were 13.6% of the attributed cost, cache writes were 45.3%, and cache reads were 40.9%.

Three levers, all free: read late, read ranges, and send bulk reads to a subagent whose context never enters the main thread.

Share my answers anonymously to improve this tool (no email, no identifiers — just the answer pattern).

Used to send this report, and occasionally something directly relevant to it. No list-selling, no sponsors, unsubscribe or ask for deletion any time. Everything else on this page stays in your browser.

For more regular thinking on this, I write essays and a practitioner digest on Substack.