> ## Documentation Index
> Fetch the complete documentation index at: https://narev.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# What Kimi K3 actually costs

> Kimi K3 lists at $3/$15 per million tokens. The route you pick matters more than the headline rate. Pricing across Moonshot, Novita, OpenRouter, and Vercel AI Gateway.

<Info>
  **In short:** Kimi K3 lists at $3 input / $15 output per million tokens. Your bill still depends on which host serves the request, whether cache reads apply, and how large the output window is. Budget the route, not just the sticker.
</Info>

Moonshot shipped [Kimi K3](https://platform.moonshot.cn) on July 16 with a 1M-token context window, multimodal input, and toggleable max-effort reasoning. The list price is simple: **$3 per million input tokens, $15 per million output tokens, \$0.30 per million cache reads.**

That number shows up on Moonshot direct, Novita, and most gateway hosts. Same model, same sticker.

The catch: "Kimi K3" isn't one API call. It's a model slug that different companies host under different routing rules, output limits, and latency profiles. Budget from the headline rate alone and you miss the bill.

This post is a running price check. Direct provider rates come from the [Narev Pricing API](https://api.narev.ai). Gateway host breakdowns are cross-checked on [OpenRouter](https://openrouter.ai/moonshotai/kimi-k3) and [Vercel AI Gateway](https://vercel.com/ai-gateway).

## Rate card

All prices are dollars per 1M tokens.

| Route                                |  Input |  Output | Cache read | Notes                                          |
| ------------------------------------ | -----: | ------: | ---------: | ---------------------------------------------- |
| Moonshot direct (`kimi-k3`)          | \$3.00 | \$15.00 |     \$0.30 | 1M context, 131K max output                    |
| Novita direct (`moonshotai/kimi-k3`) | \$3.00 | \$15.00 |     \$0.30 | 1M context, **1M max output**                  |
| OpenRouter → Moonshot                | \$3.00 | \$15.00 |     \$0.30 | \~8.8 s P50                                    |
| OpenRouter → Nebius                  | \$3.00 | \$15.00 |          — | \~0.94 s P50; no cache line on OpenRouter      |
| OpenRouter → Fireworks               | \$3.75 | \$18.75 |    \$0.375 | 25% markup over list                           |
| Vercel AI Gateway (all hosts)        | \$3.00 | \$15.00 |     \$0.30 | Flat list rate; hosts differ on speed and caps |
| Moonshot K2.7 Code (baseline)        | \$0.95 |  \$4.00 |          — | Smaller context; \~3× cheaper on input         |

Cache reads are 10% of input on Moonshot and Novita. For agent loops that reuse a large system prompt or document context, that line item matters.

Compared to K2.7 Code on Moonshot, K3 runs about **3.2× on input and 3.75× on output**. You are paying for the 1M context window and the newer stack, not a token discount.

## A real request, real money

Picture a long agent turn: 100K prompt tokens, 50K cache reads, 10K completion tokens.

| Line item              | Tokens |         Rate |         Cost |
| ---------------------- | -----: | -----------: | -----------: |
| Prompt                 |   100K |  \$3.00 / 1M |      \$0.300 |
| Cache read             |    50K |  \$0.30 / 1M |      \$0.015 |
| Completion             |    10K | \$15.00 / 1M |      \$0.150 |
| **Total (K3)**         |        |              |  **\$0.465** |
| Same turn on K2.7 Code |        |              | **\~\$0.15** |

One-third the cost, one-quarter the context. K3 isn't a drop-in replacement for K2.7 on price. It's a different tier.

## OpenRouter: one slug, three hosts

OpenRouter exposes a single model ID, `moonshotai/kimi-k3`. Behind it sit multiple inference providers. OpenRouter picks one based on your routing mode (Balanced, Nitro, Exacto). Your token rate depends on which host wins.

| Host                 | Input / Output |   Cache | P50 latency | Notes                                           |
| -------------------- | -------------- | ------: | ----------: | ----------------------------------------------- |
| Moonshot AI          | $3 / $15       |  \$0.30 |     \~8.8 s | Highest uptime of the three                     |
| Nebius Token Factory | $3 / $15       |       — |    \~0.94 s | \~9× faster than Moonshot's host; no cache line |
| Fireworks            | $3.75 / $18.75 | \$0.375 |           — | 25% markup; often highest throughput (\~46 tps) |

On OpenRouter, the slug price isn't your price. Check which host your routing mode steers toward. If you care about latency at list rates, Nebius is the interesting row. If you don't control routing, budget for the Fireworks markup as a worst case.

## Flat rates on Vercel AI Gateway

Vercel routes `moonshotai/kimi-k3` to four backends. All four list the same token rates today: $3 input, $15 output, \$0.30 cache read.

| Host        |   P50 | Throughput | Context | Max output |
| ----------- | ----: | ---------: | ------: | ---------: |
| Moonshot AI | 6.1 s |     30 tps |      1M |       131K |
| Novita AI   | 1.4 s |     19 tps |      1M |     **1M** |
| Fireworks   | 2.4 s |     69 tps |      1M |       131K |
| Baseten     | 3.1 s |     52 tps |      1M |       131K |

Fireworks on Vercel matches list price. On OpenRouter it doesn't. Same company, different gateway, different markup. Worth knowing if you switch providers and assume pricing carries over.

Novita on Vercel is the outlier on output: the only host with a **1M completion ceiling**. Every other Vercel route caps at 131K.

## Novita direct for 1M output

[Novita](https://novita.ai/models/model-detail/moonshotai-kimi-k3) hosts K3 on its own serverless API at the same $3 / $15 / \$0.30 rates. Model ID is `moonshotai/kimi-k3` (not the short `kimi-k3` slug Moonshot uses).

The reason to call Novita directly isn't price. It's the output window. Novita lists 1M max output on 1M context. Moonshot direct stops at 131K.

If you are on Vercel AI Gateway and routing to Novita anyway, going direct removes a hop and keeps the same token economics. Base URL: `https://api.novita.ai/openai`.

## Resellers and minor markups

A few third-party routes index the short `kimi-k3` slug at or near list price:

| Reseller                 | Input / Output | vs list  |
| ------------------------ | -------------- | -------- |
| Poe                      | $3.03 / $15.15 | \~1% up  |
| Venice                   | $3.75 / $18.75 | \~25% up |
| EmpirioLabs, LLM Gateway | $3 / $15       | Match    |

None of these change the story. They're convenience tax, not a different product tier.

## When K3 is worth it

* **Use K3** when your prompts routinely blow past 262K tokens (full repos, long video context, multi-hour chat history), when you need max-effort reasoning on multimodal input, or when you need Novita's 1M output ceiling.
* **Stay on K2.7 Code** when most requests fit in 32K–128K tokens and you don't need the extra context. At $0.95 / $4 on Moonshot, K2.7 is still the cost-efficient workhorse.
* **Watch your route** when buying through a gateway. OpenRouter's Fireworks host adds 25%. Vercel's doesn't. Nebius on OpenRouter matches list price with much lower latency than Moonshot's own host on the same gateway.

## Keep this current

Rates in this post get refreshed when providers move. For live numbers in your own stack:

```bash theme={null}
# Moonshot and most resellers
curl "https://api.narev.ai/v1/prices?model_id=kimi-k3&provider_id=moonshot"

# Novita (note the full model slug)
curl "https://api.narev.ai/v1/prices?model_id=moonshotai/kimi-k3&provider_id=novita"

# OpenRouter headline rate via Narev MCP
# get_gateway_prices gateway_id=openrouter model_id=moonshotai/kimi-k3
```

Gateway host breakdowns (Nebius vs Fireworks vs Moonshot on OpenRouter, per-host limits on Vercel) still live on the gateway dashboards. Narev indexes OpenRouter and Kilo as gateways; Vercel AI Gateway host rates are not in that index yet.

If you bill customers on Kimi K3 usage, wrap inference with the [Narev SDK](/docs/ai-billing/installation) and attribute cost per tenant. See [AI FinOps](/docs/guides/finops-for-ai) for the full workflow.