$1.71 to $1.02: A Long-Chat Summarisation Receipt

2026-10-26 Spend less EN

01 / The receipt: $0.69 saved across 40 more replies

A long chat can charge you repeatedly for words you have already paid to generate. Summarisation replaces part of that transcript with a shorter working record. The saving comes from sending fewer input tokens on later requests, not from making the original conversation cheaper retroactively.

Here is a worked example: keeping the full history costs $1.71 over the next 40 replies. Summarising the older history first costs $1.018, or $1.02 rounded, including the summary itself. That is $0.692 saved, approximately 40.5%.

These are illustrative calculations, not measured performance or a CheapRelay price quote. The rates below are deliberately explicit so you can substitute the actual rates for your chosen model.

02 / The assumptions behind this $1.71 bill

■ Existing conversation: 20,000 tokens, with its original generation costs already paid.
■ Next 40 exchanges: each adds a 250-token user message and a 750-token assistant reply.
■ Illustrative pricing: $1 per million input tokens and $4 per million output tokens.
■ No cache discounts, tool calls, retries, separately billed reasoning or other charges in this simplified example.
■ Both approaches produce the same assumed reply lengths; this is a budgeting assumption, not a quality guarantee.

The first new request therefore sends 20,250 input tokens. Each completed exchange adds 1,000 tokens to the next request. By request 40, the input has reached 59,250 tokens.

Across those requests, total input is 40 × 20,250 + 1,000 × 780 = 1,590,000 tokens. The 780 is the sum of zero through 39: the number of earlier exchanges replayed as the conversation grows.

03 / Receipt A: resend the complete transcript

■ Input: 1,590,000 tokens × $1 per million = $1.590.
■ Output: 30,000 tokens × $4 per million = $0.120.
■ Total for the next 40 replies: $1.710.

The output bill is small here. Repeatedly uploading the accumulating transcript is the expensive line item.

04 / Receipt B: pay $0.028 to compress the older history

Before those 40 exchanges, turn the existing 20,000-token conversation into a 2,000-token working summary. Then send that summary instead of the old transcript, while retaining every new exchange normally.

■ Summary input: 20,000 tokens × $1 per million = $0.020.
■ Summary output: 2,000 tokens × $4 per million = $0.008.
■ Later chat input: 870,000 tokens × $1 per million = $0.870.
■ Later chat output: 30,000 tokens × $4 per million = $0.120.
■ Total, including compression: $1.018.

The replacement removes 18,000 input tokens from each subsequent request: 720,000 tokens across 40 requests. That avoids $0.720 of input charges, less the $0.028 summary fee. Under these assumptions, compression pays for itself after two further requests.

05 / Keep decisions and identifiers, not conversational padding

A useful working summary preserves the current goal, hard constraints, decisions, unresolved questions and exact identifiers. Keep critical code, error messages or contractual wording verbatim when paraphrasing could change their meaning. Retain the latest exchanges separately so immediate context remains intact.

Keep the original transcript outside the active prompt for retrieval. If a later question depends on discarded detail, restore the relevant passage. This example compresses only once; the new history still grows. Repeated compaction needs its own cost and accuracy checks.

06 / Check cached-token charges before claiming the saving

Cached history can be cheaper than ordinary input. Replacing it with a summary may also change cache reuse. Compare actual input, output, cache-read and cache-write charges rather than applying this uncached receipt blindly.

CheapRelay bills per real token, including cache reads and writes, with no monthly fee or subscription. Sign up for CheapRelay to test a small workload, then compare your itemised usage before and after summarisation. Use the measured bill—not the shortest-looking prompt—as the final verdict.

More reading

Date	Aisle	Written in	Title
2026-10-26	Spend less	EN	$30 a Month for AI: An Indie Hacker’s Receipt A $30 monthly AI budget with token math, cache allowances, retry limits, and practical spending controls for solo builders.
2026-10-26	Spend less	EN	$3.45 for 95 accepted answers beats cheap tokens A worked receipt shows why token prices can mislead: count retries, rejected outputs and cache charges before ranking models.
2026-10-26	Spend less	EN	The $0.004 AI Task With a $0.0566 Agent Receipt A hypothetical agent receipt shows how context, retries, evaluation and logging turn a $0.004 task into $0.0566 per accepted result.
2026-10-26	Spend less	EN	$2.53 for 10,000 JSON Jobs: Read the Receipt A worked JSON-output budget: token costs, schema overhead, retries, and the checks that keep cheap API calls from becoming expensive workflows.
2026-10-26	Spend less	EN	$0.012 Either Way: Streaming Changes the Habit A token-by-token receipt shows why streaming has no automatic discount, but stopping early and avoiding retries can change your API bill.
2026-10-26	Spend less	EN	¥0.32 for a 100,000-token translation batch A worked translation receipt shows how to rank model quotes by yuan spent, then adjust for retries, cache charges and human review.