Updated 2026-10-27
Kimi K3 Pricing Explained
Kimi K3 costs ¥20/MTok input and ¥100/MTok output. See K3.1 Pro, highspeed, cache pricing, and a worked monthly workload example.
TL;DR
- Kimi K3 list price: ¥20 per 1M input tokens and ¥100 per 1M output tokens.
- Kimi K3.1 Pro list price: ¥13/MTok input and ¥68/MTok output; the highspeed version is ¥26/MTok input and ¥136/MTok output.
- There is no flagship free tier: K3 and K3.1 Pro require a minimum ¥10 balance, and the new-user ¥15 voucher does not apply to them.
- A 50-chat/day workload using 2,000 input and 500 output tokens per chat costs about ¥135/month with Kimi K3 and ¥90/month with K3.1 Pro before cache savings.
Kimi K3 pricing table
| Model | Input / MTok | Output / MTok | Cache hit / MTok | Notes |
|---|---|---|---|---|
| kimi-k3 | ¥20 | ¥100 | ¥2 | Previous flagship; weights released |
| kimi-k3.1-pro | ¥13 | ¥68 | ¥1.30 | Current flagship; 1M context |
| kimi-k3.1-pro-highspeed | ¥26 | ¥136 | ¥1.30 | About 2–3× faster output |
What a real workload costs
Assume 50 chats per day for 30 days. Each request sends 2,000 input tokens and receives 500 output tokens. Monthly usage is 3.0M input tokens and 0.75M output tokens.
With Kimi K3: 3 × ¥20 + 0.75 × ¥100 = ¥60 + ¥75 = ¥135 per month. With K3.1 Pro: 3 × ¥13 + 0.75 × ¥68 = ¥39 + ¥51 = ¥90 per month. Doubling traffic to 100 chats/day approximately doubles those totals to ¥270 and ¥180.
If the same knowledge base or system prompt is repeated, context caching can reduce the input cost dramatically: cache-hit input is ¥2/MTok for K3 and ¥1.30/MTok for K3.1 Pro. Cache writes add a separate one-time charge.
Cache and hidden costs to watch
| Cost type | Kimi K3 | Kimi K3.1 Pro | Highspeed K3.1 Pro |
|---|---|---|---|
| Cache hit / MTok | ¥2 | ¥1.30 | ¥1.30 |
| Cache write, 5-minute TTL / MTok | ¥20 | ¥13 | ¥26 |
| Cache write, 1-hour TTL / MTok | ¥40 | ¥26 | ¥52 |
Why bills can be higher than expected
- Reasoning models can consume output tokens before the visible answer is produced; cap max_completion_tokens when cost matters.
- Agent loops and failed tool calls repeat context and generate additional output tokens.
- Every document or file excerpt passed as input is billed like normal prompt text, even if file extraction itself is currently free.
- A cache miss or a partial cache block cannot receive the cache-hit price, so short or frequently changed prefixes may not save money.
Is Kimi K3 worth the price?
For long-context coding and agent workflows, K3.1 Pro is unusually inexpensive at its published list price: it is cheaper than K3 while offering stronger published benchmark results. The highspeed variant doubles input and output list prices in exchange for approximately 2–3× faster output.
For simple classification, extraction, routing, or very high-volume support automation, a smaller model may be more cost-effective. Benchmark your real prompts and measure token usage before committing.
Frequently asked questions
How much does Kimi K3 cost?
As of 2026-10-27, Kimi K3 costs ¥20 per 1M input tokens and ¥100 per 1M output tokens. Cached input costs ¥2/MTok when it hits.
Is Kimi K3 free to use?
No. Moonshot says flagship K3 and K3.1 Pro access is unlocked after a minimum ¥10 top-up. The ¥15 new-user voucher cannot be used for these flagship models.
Why is my Kimi K3 bill higher than expected?
The usual causes are hidden reasoning output, repeated long context in agent loops, cache misses, retry storms, and high output limits. Inspect token usage per request and use caching for repeated prefixes.