Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens on Moonshot's official API, with cached input at $0.30. That price has not moved since our first check in August. The market around it has. Twenty providers now host K3 across 24 listings, and the same model lists from $0.10 to $4.50 per million input tokens and from $4.00 to $22.50 per million output tokens.
When we first documented K3 on August 27, 2026, eleven hosts priced it within 1.23x of each other, the tightest market we had seen for an open-weight model. Six weeks later the input spread is 45x. This page has the official rates, every host, what a real workload pays, and which number in the table actually decides the bill.
What Kimi K3 is, briefly
Kimi K3 is Moonshot AI's flagship, a 2.8 trillion parameter open-weight multimodal reasoning model with a 1M token context window, vision, tool calling, and prompt caching. The API launched July 16, 2026, and the weights followed on July 27 under a modified MIT license.
On Artificial Analysis benchmarks, as shown on OpenRouter, K3 scores better than 93 percent of models compared on intelligence, 90 percent on coding, and 86 percent on agentic tasks. It is priced as a premium reasoning model, and on the official API it is still the most expensive open-weight flagship we track.
Kimi K3 official API pricing
Rates from Moonshot's pricing page, USD per 1M tokens:
| Token type | Price per 1M tokens |
|---|---|
| Input | $3.00 |
| Cached input | $0.30 |
| Output | $15.00 |
| Cache write, 5 minute TTL | $3.00 |
| Cache write, 1 hour TTL | $6.00 |
Pricing is flat across the full 1,048,576 token window with no long-context surcharge. Cached input is billed at a 90 percent discount. Cache writes are billed separately by how long the entry lives: the default 5 minute tier costs the same as normal input, and the 1 hour tier costs double.
At $15 per million output tokens, K3's official rate is about 3.8 times the peak output rate of DeepSeek V4 Pro and about 3.6 times the original output price of Mistral Large 4.
Kimi K3 pricing by provider
All listings on OpenRouter's standard routing as of October 8, 2026, sorted by input price. USD per 1M tokens. Speed is median tokens per second.
| Provider | Input | Output | Cached input | Speed |
|---|---|---|---|---|
| Relace | $0.10 | $4.00 | $0.10 | 85 tps |
| inference.net | $0.53 | $12.00 | $0.30 | 68 tps |
| Wafer | $0.79 | $15.00 | $0.45 | 44 tps |
| Sail Research | $0.84 | $13.50 | $0.30 | 60 tps |
| AkashML | $1.20 | $14.00 | $1.20 | 39 tps |
| Makora | $1.53 | $12.75 | $0.204 | 31 tps |
| Morph | $1.676 | $14.00 | $0.29 | 64 tps |
| Decart (35% off promotion) | $1.95 | $9.75 | $0.195 | 6 tps |
| DigitalOcean | $2.55 | $12.95 | $0.255 | 36 tps |
| Phala (15% off promotion) | $2.55 | $12.75 | $0.255 | 48 tps |
| Parasail | $2.60 | $13.00 | $0.26 | 66 tps |
| Together | $2.70 | $13.50 | $0.27 | 43 tps |
| Wafer (second listing) | $2.80 | $14.00 | $0.30 | 43 tps |
| DeepInfra | $2.85 | $14.25 | $0.285 | 43 tps |
| inference.net Fast | $2.99 | $15.00 | $0.45 | 80 tps |
| Parasail Fast | $3.00 | $15.00 | $0.30 | 46 tps |
| Amazon Bedrock | $3.00 | $15.00 | $0.30 | 60 tps |
| Chutes | $3.00 | $15.00 | $0.30 | 17 tps |
| Modal | $3.00 | $15.00 | $0.30 | 63 tps |
| Fireworks | $3.00 | $15.00 | $0.30 | 26 tps |
| Baseten | $3.00 | $15.00 | $0.30 | 74 tps |
| Moonshot AI (official) | $3.00 | $15.00 | $0.30 | 21 tps |
| Alibaba Cloud International | $3.45 | $17.25 | $0.345 | 37 tps |
| Fireworks Fast | $4.50 | $22.50 | $0.45 | 75 tps |
Three things to read from this table.
The spread is on input, not output. Input runs from $0.10 to $4.50, a 45x spread. Output runs from $4.00 to $22.50, and if you set aside the single lowest listing, from $9.75 to $22.50. Twenty of the 24 listings charge between $12 and $15 for output.
The official API is no longer the reference price. Fifteen listings sit below Moonshot's $3.00 input, seven match it, and two charge more. Moonshot is also one of the slower hosts at 21 tokens per second.
One listing is an outlier. Relace at $0.10 and $4.00 is far below every other host. Before routing volume there, verify precision, context limits, and rate limits. A price that far from the field is a reason to test, not a reason to assume.
From 1.2x to 45x in six weeks
| August 27, 2026 | October 8, 2026 | |
|---|---|---|
| Providers | 11 | 20 (24 listings) |
| Input range | $2.80 to $3.45 | $0.10 to $4.50 |
| Output range | $14.00 to $17.25 | $4.00 to $22.50 |
| Input spread | 1.23x | 45x |
| Official price | $3.00 / $15.00 | $3.00 / $15.00 |
In August we wrote that six weeks was early and that every market we had tracked had moved. This one moved too. The official price stayed where it was, more hosts picked up the open weights, and the single price became two dozen.
The pattern in how they priced it is worth noticing. Hosts competed on the input rate and mostly left output alone. OpenRouter's own usage data explains why. Across K3 traffic the weighted average price actually paid is $0.45 per million input tokens and $13.93 per million output tokens, because cache hit rates on most hosts run around 90 percent. Buyers were already paying close to the cached rate on input. A low headline input price costs the host little, and output is where the revenue is.
That is the same reason one model can carry many prices at once, covered in why token prices differ.
What a real workload pays
A production assistant doing 10M input and 1M output tokens per day, with no caching:
| Provider | Daily | Annual |
|---|---|---|
| Relace ($0.10 / $4.00) | $5.00 | $1,825 |
| inference.net ($0.53 / $12.00) | $17.30 | $6,315 |
| Moonshot official ($3.00 / $15.00) | $45.00 | $16,425 |
| Fireworks Fast ($4.50 / $22.50) | $67.50 | $24,638 |
Same model, same tokens, and the annual bill runs from under $2,000 to over $24,000 depending on the listing.
Caching changes the picture. With a 90 percent cache hit rate on input, the official API drops to $20.70 a day, about $7,556 a year. The same workload on inference.net drops to $15.23 a day, about $5,559 a year. Without caching, the official API costs 2.6 times what inference.net does. With heavy caching it costs about 1.4 times as much.
So for cache-heavy agent workloads, the headline input price matters much less than it looks. The output rate and the cached input rate decide the bill, and on those two numbers most hosts are close together.
For comparison, the same uncached workload on DeepSeek V4 Pro off-peak runs about $3,132 a year. See the LLM API pricing comparison for the cross-model view.
How to choose a K3 provider
- Compare output and cached input first. Those are the rates most of your bill runs on. A host with a low input price and a $15 output price is not much cheaper for a cached workload.
- Check speed alongside price. Median throughput runs from 6 to 85 tokens per second across hosts. Some of the cheapest listings are also among the fastest, and one promotional listing is the slowest on the table.
- Treat promotions as temporary. Two listings carry a marked discount (35 percent and 15 percent off). Budget at the list price behind them.
- Verify the outlier. Test any listing far below the field for precision, context length, and rate limits before moving production traffic.
- Lock what you can. K3 went from one price to two dozen in six weeks, and the listings will keep moving. For predictable volume, a forward contract fixes capacity at a known price. See how to lock in token prices and the Mercatus Forward Market.
To track token prices across models and hosts, see the Mercatus Token Index.
Frequently asked questions
How much does Kimi K3 cost per million tokens?
$3.00 input and $15.00 output on Moonshot's official API, with cached input at $0.30. Third-party listings range from $0.10 to $4.50 input and $4.00 to $22.50 output as of October 8, 2026.
Who is the cheapest Kimi K3 provider?
Relace lists the lowest price at $0.10 input and $4.00 output, well below every other host. The next lowest is inference.net at $0.53 input and $12.00 output. Verify precision and limits before relying on the lowest listing.
How many providers host Kimi K3?
Twenty providers across 24 listings on OpenRouter's standard routing as of October 8, 2026, up from eleven on August 27, 2026.
Does Kimi K3 charge for cache writes?
Yes. On the official API, cache writes cost $3.00 per 1M tokens at the default 5 minute TTL and $6.00 at the 1 hour TTL. Cached input that hits within the entry's lifetime is billed at $0.30 with no extra write charge.
Is Kimi K3 open weight?
Yes. Moonshot released the weights on July 27, 2026 under a modified MIT license, eleven days after the API launch.
Does Kimi K3 charge more for long context?
No. Flat pricing applies across the full 1M token window on the official API.
Why do Kimi K3 prices differ so much between providers?
Hosts run the same open weights on different hardware, at different utilization, with different caching and margin targets. Most compete on the input rate and keep output near the official $15, because cached traffic already pays little for input.
Methodology
Official rates and cache write terms from Moonshot's Kimi API pricing page. Provider listings, throughput, weighted average paid prices, and cache hit rates from OpenRouter's Kimi K3 page, standard routing, read October 8, 2026. One additional Fireworks listing that OpenRouter does not use in standard routing is excluded. Promotional listings are shown at the discounted rate with the discount noted. Benchmark percentiles are Artificial Analysis figures as displayed by OpenRouter. August 27, 2026 figures are from the earlier version of this page. Workload examples use stated token volumes at 365 days and ignore cache write charges; recompute with your own mix. Last verified: 2026-10-08.