Kimi K3 vs Claude Sonnet: The Long-Document Receipt

01 / A 200,000-Token Document Needs an Itemized Bill

Comparing Kimi K3 with Claude Sonnet 5.5 for long documents starts with four receipt lines: uncached input, cache writes, cache reads, and output. A cheap-looking input rate is not the whole purchase.

The examples below use published provider rates checked on October 25, 2026, not verified CheapRelay prices. They are calculated budgets, not measured performance results. CheapRelay’s current route prices and actual usage determine your relay bill.

For an initial input-and-output comparison, Kimi K3 is listed at $3 per million uncached input tokens and $15 per million output tokens. Claude Sonnet 5.5 lists $2 and $10 respectively. Cache operations require separate treatment. ([tomshardware.com](https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3?utm_source=openai))

02 / Receipt One: One Long Report, One Short Answer

Assume each model records exactly 200,000 uncached input tokens, including the document and instructions, and 6,000 billable output tokens. Equal token counts isolate pricing; they do not mean the same file necessarily produces identical counts.

• ■ Kimi K3 input: 200,000 / 1,000,000 × $3 = $0.60.
• ■ Kimi K3 output: 6,000 / 1,000,000 × $15 = $0.09.
• ■ Kimi K3 input-and-output subtotal: $0.69.
• ■ Claude Sonnet 5.5 input: 200,000 / 1,000,000 × $2 = $0.40.
• ■ Claude Sonnet 5.5 output: 6,000 / 1,000,000 × $10 = $0.06.
• ■ Claude Sonnet 5.5 input-and-output subtotal: $0.46.

Using those published rates, the subtotal difference is $0.23 per document, or $23 across 100 documents. These are not complete invoices: applicable cache writes, tools, extraction services, taxes, and retries remain outside this calculation. Kimi’s documentation explicitly separates cache-write billing, so do not silently treat its $0.69 subtotal as an all-inclusive total. ([tomshardware.com](https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3?utm_source=openai))

03 / Receipt Two: Nine Follow-Ups Against a Warm Cache

Now suppose the 200,000-token document is already cached. Ask nine independent follow-up questions, each adding 1,000 fresh input tokens and generating 2,000 billable output tokens. Assume all nine document reads hit the cache; exclude earlier setup costs and conversation history.

Kimi K3 publishes a $0.30-per-million cache-read rate. Claude Sonnet 5.5 publishes $0.10 per million. Those rates make the repeated-document portion straightforward to price. ([platform.kimi.ai](https://platform.kimi.ai/docs/guide/context-caching))

• ■ Kimi K3 cached document reads: 1,800,000 tokens × $0.30 / 1,000,000 = $0.540.
• ■ Kimi K3 fresh questions and output: $0.027 + $0.270 = $0.297.
• ■ Kimi K3 nine-follow-up subtotal: $0.837.
• ■ Claude Sonnet 5.5 cached document reads: 1,800,000 tokens × $0.10 / 1,000,000 = $0.180.
• ■ Claude Sonnet 5.5 fresh questions and output: $0.018 + $0.180 = $0.198.
• ■ Claude Sonnet 5.5 nine-follow-up subtotal: $0.378.

The calculated difference is $0.459 for these follow-ups. This receipt deliberately starts after cache creation: add the initial request and write charges before comparing complete sessions.

04 / The Cache-Write Line Can Change Your Budget

Kimi K3 lists cache writes at $3 per million tokens for five minutes and $6 for one hour. A 200,000-token write therefore represents $0.60 or $1.20. Claude Sonnet 5.5 lists $2.50 and $4, giving write charges of $0.50 or $0.80. Apply each route’s billing semantics rather than assuming write and uncached-input categories work identically. ([platform.kimi.ai](https://platform.kimi.ai/docs/guide/context-caching))

For repeated document questions, keep reference text unchanged at the beginning and changing questions at the end. Kimi’s caching documentation explains that prefix changes prevent reuse downstream; its default cache lifetime is five minutes. Verify actual hits instead of budgeting every repeat as cached. ([platform.kimi.ai](https://platform.kimi.ai/docs/guide/context-caching))

05 / Choose by Cost per Accepted Document

Run a small acceptance test before committing volume: ten documents, identical questions, and a checklist covering omissions, quotation accuracy, and unsupported claims. Record each route’s actual tokens and total charges. Divide spending by accepted answers, not merely requests sent. The receipts above establish price arithmetic, not a quality winner.

CheapRelay offers pay-as-you-go access through one key, with no monthly fee or subscription and billing for real tokens, including cache reads and writes. That makes a small comparison practical without assuming either route’s price matches its provider list rate.

Sign up for CheapRelay and compare both models on your own documents, then keep the route with the lower verified cost per usable result.