Kimi K3 API: Pricing, Context Window and Setup

Kimi K3 is Moonshot AI's flagship reasoning model for software engineering, long-horizon agent work, visual understanding and large knowledge tasks. The API model ID is kimi-k3, the context window is 1,048,576 tokens, and the official standard price is $3 per 1M cache-miss input tokens and $15 per 1M output tokens.

Kimi K3 is now available on RouterPlex through the same OpenAI-compatible endpoint used for GPT, Claude, Gemini and the rest of the catalog.

Kimi K3 API pricing #

Moonshot defines 1M as 1,000,000 tokens for billing. Cache hits cost one tenth of normal input, so repeated long instructions, tool definitions and conversation prefixes can be materially cheaper when the provider reuses the cached prefix.

RouterPlex bills those same Kimi K3 categories at the published rates. A $5 prepaid balance therefore buys roughly 1.67 million cache-miss input tokens, 16.67 million cached input tokens, or 333,333 output tokens if the workload used only one category. Real requests combine input and output.

See the live Kimi K3 API price page for the current RouterPlex catalog rate.

Kimi K3 pricing on OpenRouter vs RouterPlex #

Per-token prices for Kimi K3 are the same on OpenRouter and RouterPlex, because both pass through Moonshot's published rates. The cost difference is the credit purchase fee. Buying $100 of Kimi K3 credit costs $105.50 by card on OpenRouter, or $100 on RouterPlex.

The per-token rate is what most comparisons publish, so the two look identical on a price page. The fee applies whenever credit is purchased, which means it scales with how much credit you buy rather than how many tokens you burn. On small top-ups the $0.80 card minimum dominates: $5 of credit costs $5.80, an effective 16%. For the full breakdown, see OpenRouter credits, fees and expiry and the RouterPlex vs OpenRouter comparison.

What Kimi K3 costs on a real workload #

List prices per 1M tokens are hard to reason about. Here is a single agent turn that sends a 40K-token prompt and returns 2K tokens of output, priced three ways depending on how much of the prompt hits Moonshot's cache:

The arithmetic for the 75% row:

30,000 cached tokens   × $0.30/1M  = $0.009
10,000 uncached tokens × $3.00/1M  = $0.030
 2,000 output tokens   × $15.00/1M = $0.030
                              total = $0.069

Two things fall out of this. First, cache design matters more than model choice on Kimi K3 — a well-structured stable prefix cuts the input bill by two thirds, which dwarfs most price differences between comparable models. Keep system prompts, tool definitions and file context in a fixed prefix so the cacheable portion stays byte-identical between turns.

Second, output becomes the dominant cost once caching works. Output is 20% of an uncached turn but 57% of a 90%-cached one. Reasoning models emit a lot of tokens, so once the cache is working, capping max_tokens is a more effective lever than shaving the prompt.

Is there a Kimi K3 subscription? #

No. Kimi K3 API access is billed per token, not by subscription. Kimi's consumer subscription plans cover the Kimi chat app and do not include API keys, so there is no monthly seat cost to add on top of token spend. A prepaid balance is the entire cost of API access.

That matters for budgeting: with no recurring floor, a Kimi K3 workload that runs for one week costs one week of tokens. RouterPlex follows the same model — a prepaid balance, per-key hard budgets, and no subscription tier gating model access.

Kimi K3 specifications #

Moonshot describes Kimi K3 as a 2.8-trillion-parameter model built with Kimi Delta Attention and Attention Residuals. Those architecture details matter less operationally than the API behavior: K3 can retain a large working context, inspect images, reason before answering and participate in tool-calling loops.

Call Kimi K3 with RouterPlex #

Create a RouterPlex key, give it a hard budget, and send the standard chat-completions request:

curl https://api.routerplex.com/v1/chat/completions   -H "Authorization: Bearer $ROUTERPLEX_API_KEY"   -H "Content-Type: application/json"   -d '{
    "model": "kimi-k3",
    "messages": [
      {"role": "user", "content": "Review this migration plan and identify the highest-risk assumption."}
    ],
    "reasoning_effort": "max"
  }'

The equivalent Python setup uses the regular OpenAI client:

import os
from openai import OpenAI
 
client = OpenAI(
    api_key=os.environ["ROUTERPLEX_API_KEY"],
    base_url="https://api.routerplex.com/v1",
)
 
response = client.chat.completions.create(
    model="kimi-k3",
    reasoning_effort="max",
    messages=[{"role": "user", "content": "Design a rollback-safe deployment plan."}],
)
 
print(response.choices[0].message.content)

Where Kimi K3 fits #

Kimi K3 is most relevant when a task needs more than a short answer:

• Repository-scale planning with many files and constraints.
• Long-running agents that call multiple tools.
• Large document collections or extended conversation state.
• Visual inspection combined with technical reasoning.
• Structured JSON output for downstream automation.

For a dedicated coding model with a smaller 256K context, Kimi K2.7 Code may still be a useful comparison. Kimi K3 is the broader flagship choice when coding is mixed with research, planning, images or long-context knowledge work.

Kimi K3 API details to watch #

K3 always reasons. Moonshot currently accepts only reasoning_effort="max", although more levels are planned. Streaming responses can expose reasoning and final-answer deltas separately.

For multi-turn tool use, retain the complete assistant message returned by the model. Keeping only the visible text can remove tool-call or reasoning state needed by the next request. Moonshot also notes that its web-search capability is being updated, so do not make a production design depend on that feature without retesting it.

Run a controlled Kimi K3 test #

Start with one narrow workload and a hard key budget. Compare answer quality, latency, cached-token behavior and total cost against the model you currently use.

Create a RouterPlex account, add the minimum $5 balance, and select kimi-k3. You can compare it with GPT-5.6 and Claude Fable 5 through the same endpoint.