$0.60 a turn: the 200k-context receipt

01 / A nearly full 200k window costs $0.60 in this example

A 200,000-token context window is capacity, not a flat-rate purchase. The bill comes from the tokens processed on each request, their billing categories, and the applicable rates. A large window becomes expensive when an application keeps sending the same large conversation back.

Start with an illustrative rate card: $3 per million uncached input tokens and $15 per million output tokens. These are calculation assumptions, not a quoted CheapRelay price. Check the selected model’s current rates before budgeting.

Our request contains 190,000 input tokens and produces 2,000 output tokens. That leaves room within a nominal 200,000-token combined context limit. Actual context accounting and output limits depend on the model.

• ■ Input: 190,000 ÷ 1,000,000 × $3 = $0.57.
• ■ Output: 2,000 ÷ 1,000,000 × $15 = $0.03.
• ■ Total receipt: $0.60 for one successful request.

The unused capacity does not itself create a token charge. However, system instructions, tool definitions, retrieved documents, and conversation history can all occupy input space. Count the complete request, not just the text someone pasted.

02 / Twenty turns turn $0.60 into $12

Suppose an agent makes 20 requests, each with 190,000 input tokens and 2,000 output tokens. Assume no caching and a managed history that keeps input size constant.

• ■ Total input: 3,800,000 tokens, costing $11.40.
• ■ Total output: 40,000 tokens, costing $0.60.
• ■ One 20-request session: $12.00.
• ■ Ten such sessions per working day: $120.
• ■ Twenty working days: $2,400.

This is a workload projection, not a usage benchmark. Real conversations may grow, shrink, stop early, or invoke additional tools. The important point is that earlier context can be billed again when it is submitted again. A chat interface showing one conversation does not imply one input charge.

03 / A reusable 180k prefix cuts this receipt to $2.90

Now split each request into a stable 180,000-token prefix and 10,000 fresh input tokens. Keep the same 2,000-token output. For illustration, assume cache writes cost $3.75 per million tokens and cache reads cost $0.30 per million. Again, these are hypothetical rates, not a live offer.

The first request writes the stable prefix. The next 19 requests successfully reuse it:

• ■ First cache write: 180,000 ÷ 1,000,000 × $3.75 = $0.675.
• ■ Fresh input per request: 10,000 ÷ 1,000,000 × $3 = $0.03.
• ■ Output per request: $0.03.
• ■ First request total: $0.735.
• ■ Each later cache read: 180,000 ÷ 1,000,000 × $0.30 = $0.054.
• ■ Each later request total: $0.114.
• ■ Session total: $0.735 + 19 × $0.114 = $2.901, or $2.90 rounded.

Against the uncached $12 receipt, that saves approximately $9.10. The calculation assumes every later prefix reuse qualifies for a cache hit. Expiration, changed content, minimum cache sizes, and provider-specific rules can change the result. Cache writes are not free, and a mostly unique prompt may gain little.

04 / Retries and unnecessary context add separate receipt lines

One additional fully processed, uncached request adds another $0.60 under our assumptions. A failed request may have different billing treatment; do not assume every failure is free or every retry costs the full amount.

Removing 50,000 unnecessary uncached input tokens saves $0.15 per request at the example rate. Across 20 requests, that is $3. Smaller, relevant inputs can be worth more than buying access to a larger window.

05 / Compare the billable categories before funding usage

CheapRelay offers pay-as-you-go API access with no monthly fee or subscription, billing per real token, including cache reads and writes. That makes input, output, and cache rates the useful comparison—not context capacity alone.

Sign up for CheapRelay and test a representative session with a small budget. Record actual input, output, cache activity, and total charges. Your best cost estimate is a measured receipt multiplied by expected workload, with room for retries—not the largest token number on a model listing.