kimik3.io/Pricing/Calculator

Kimi K3 cost calculator

Your workload, the official rates, and what prompt caching actually saves. Everything runs in your browser — nothing is sent anywhere.

Your workload

System prompt + tool schemas + documents — the part that repeats byte-identically. This is the only part that can cache.

The user turn — always billed at the full $3.00/1M.

Includes reasoning — we measured it at 43–77% of output. Budget several times your answer length.

Share of prefix tokens served from cache. Steady agent loops run high; anything that drifts the prefix pulls it down.

Estimated monthly bill · 30 days

$—
Prefix input —
Varying input at $3.00/1M—
Output at $15.00/1M, reasoning included—
Total—
Same workload, zero cache hits—
Caching saves you—

Rates from the official pricing page, read 2026-07-16 — the same rates you pay through EvoLink. This is an estimate, not a quote: the first call of every cold period bills at full price, up to 255 tokens of tail never cache, and your real reasoning share varies by prompt.

Three reference workloads

The same math as the calculator above, precomputed — for orientation, and for readers without JavaScript. Official 2026-07-16 rates; 30-day months.

Monthly cost = prefix×(hit·$0.30 + miss·$3.00) + varying×$3.00 + output×$15.00, per 1M tokens.
WorkloadShapeMonthlyWithout caching
Light chatbot 1,000 req/day · 400 varying input · 600 output · no repeated prefix $306 $306 (nothing repeats)
Agent loop 200 req/day · 50k fixed prefix at 90% hit · 500 varying · 800 output $252 $981
Long-document Q&A 20 req/day · 400k document at 95% hit · 200 varying · 1,000 output $114 $729

The chatbot row is an output-cost problem — caching can't help a prompt that never repeats, and reasoning is inside that output number. The other two rows are why prefix layout is worth an afternoon: same traffic, 4–6× apart.

How the math works

Input tokens split into two streams. The fixed prefix bills at $0.30/1M when it hits the cache and $3.00/1M when it misses — your hit rate blends the two. The varying turn can never cache, so it always bills at $3.00/1M. Output bills at $15.00/1M, and on K3 that includes the reasoning tokens you never see — which is why the output field above deserves a bigger number than your answers alone would suggest.

The single biggest lever is the hit rate — drag the slider from 0 to 90% and watch the prefix line. Structuring your prompt to keep it high is a layout problem, not a spend problem: stable bytes first, varying turn last, no drift.

Now run it for real

EvoLink carries kimi-k3 at these same rates on an OpenAI-compatible endpoint — 10 free credits, sign up from anywhere — no Chinese phone number.

Get an EvoLink API key