Moonshot AI's Kimi K3 has dominated model-release headlines this month: the largest open-weight model shipped in 2026, with benchmark results rivaling the top closed American systems. Searches for "Kimi K3 API" have followed — here is the pricing, the whole K-series lineup, and how to get access without a Chinese phone number.
| Model | Input / 1M | Output / 1M | Positioning |
|---|---|---|---|
| kimi-k2.5 | $0.638 | $3.348 | Budget workhorse |
| kimi-k2.6 | $1.036 | $4.305 | General flagship |
| kimi-k2.7-code | $1.036 | $4.305 | Code-specialized |
| kimi-k2.7-code-highspeed | $2.073 | $8.610 | Low-latency coding |
| kimi-k3 | $2.90 | $14.50 | Frontier open-weight |
All prices are official-direct, USD-billed, and include the full 1M-token context window. No platform markup, no subscription.
from openai import OpenAI
client = OpenAI(api_key="sk-novai-...", base_url="https://aiapi-pro.com/v1")
r = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Review this patch for race conditions."}],
)
print(r.choices[0].message.content)
One quirk worth knowing: the Kimi family is served with a fixed sampling temperature upstream — the proxy pins it for you, so don't waste time debugging a temperature parameter that appears ignored. It is expected behavior, not a bug.
kimi-k3 for frontier-quality reasoning and long-document work where accuracy beats cost. kimi-k2.7-code remains the value pick for agentic coding pipelines at $1.036/$4.305, and the highspeed variant is for interactive tools where latency is user-facing. If your workload is high-volume classification or extraction, kimi-k2.5 at $0.638 input is hard to beat at this context length.
Rule of thumb: prototype on kimi-k2.6, ship on the cheapest model that passes your eval set, and reserve kimi-k3 for the tasks where nothing else passes.
Moonshot's own platform requires a Chinese phone number and CNY billing. NovAI proxies the official-direct endpoint with international payment (USDT (TRC20)), and every new account gets $2 of free credit — enough to send roughly half a million input tokens through kimi-k3 before paying anything.
Related: Best coding API of 2026: Kimi vs DeepSeek vs Qwen · The cheapest LLM APIs this month