Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens, with cached input at $0.30 per million. Those three numbers are the whole headline, and the third one is the interesting one — a 10x discount on repeated context is what makes long-context work on this model economically different from its competitors.
The full price list
| Mode | Price per 1M tokens |
|---|---|
| Input (cache miss) | $3.00 |
| Input (cache hit) | $0.30 |
| Output | $15.00 |
Automatic context caching is supported, so the cache-hit rate is not something you have to build infrastructure for — you get it by keeping your prompt prefix stable. Prices exclude applicable taxes, which are calculated at checkout by jurisdiction.
How it compares
| Model | Input | Output |
|---|---|---|
| Claude Fable 5 | $10 | $50 |
| GPT-5.6 Sol | $5 | $30 |
| Claude Opus 5 | $5 | $25 |
| Kimi K3 | $3.00 | $15.00 |
| Claude Sonnet 5 | $3 | $15 |
| Claude Haiku 4.5 | $1 | $5 |
K3 lands exactly on Claude Sonnet 5's rate — a mid-tier price for a frontier-scale model. Against Claude Opus 5 it is 40% cheaper; against Claude Fable 5 it is roughly a third of the cost on output. Competitor figures are consistent with those used across the rest of our coverage.
What that costs in real requests
Token prices are abstract until you convert them. Rough guide, at standard uncached rates:
| Task | Input | Output | Approximate cost |
|---|---|---|---|
| A short chat exchange | 1K | 1K | Under $0.02 |
| Reviewing a 5,000-line diff | 60K | 3K | ~$0.23 |
| Analysing a 200-page report | 150K | 4K | ~$0.51 |
| A full 1M-token prompt | 1M | 8K | ~$3.12 |
| A 100-step agent run | 400K | 60K | ~$2.10 |
Every figure above falls by up to 90% on the input side once your stable prefix is hitting cache. That is the single biggest lever on this model.
The reasoning cost you cannot switch off
One structural point that pricing tables hide: reasoning is always on. Every request pays for a reasoning pass, and those tokens bill as output at $15.00 per million. There is no configuration that removes this.
In practice this means K3 is comparatively expensive for trivial high-volume calls and comparatively cheap for hard ones. If your workload is a million short classifications a day, a smaller non-reasoning model will beat it on cost regardless of headline rates. If your workload is a thousand genuinely difficult tasks, K3 is priced well below what that capability usually costs.
Every way to pay less
- Cache the stable part of your prompt. $0.30 against $3.00 is a 10x saving on input. Put fixed context first and vary only the tail.
- Drop
reasoning_efforttolowon mechanical work. The default ismax, and the default is wrong for a large share of real traffic. - Ask for shorter output. Output bills at 5x input. A conciseness instruction is a direct discount.
- Route by difficulty. Send easy work to a cheaper model entirely and reserve K3 for tasks that need it.
- Don't fill the window out of habit. A large prompt is a real cost with no quality guarantee.
More on Kimi K3
Start with the complete Kimi K3 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Kimi K3 guideFrequently Asked Questions
How much does Kimi K3 cost?
$3.00 per million input tokens and $15.00 per million output tokens. Cached input drops to $0.30 per million — a 10x discount that applies automatically when your prompt prefix stays stable between requests.
Is Kimi K3 cheaper than Claude or GPT?
Cheaper than the flagship tiers: 40% below Claude Opus 5's $5/$25, half the output price of GPT-5.6 Sol's $5/$30, and about a third of Claude Fable 5's $10/$50. It matches Claude Sonnet 5 exactly at $3/$15, and it is more expensive than small models like Claude Haiku 4.5 at $1/$5.
What is the cheapest way to run Kimi K3?
Cache aggressively and lower the effort level. Cached input is a tenth of the standard rate, and `reasoning_effort: low` cuts the reasoning tokens that bill as output. Beyond that, route easy requests to a cheaper model — with reasoning always on, K3 is not the right tool for high-volume trivial work.
