# Kimi K3 Pricing and Cost Explained for Operators

> Canonical: https://www.yalc.ai/blog/kimi-k3-pricing-cost/

Published API rates, cost versus Claude and Codex, and worked examples so you know what a month of agent work actually costs.

Kimi K3 pricing is 3.00 dollars per million input tokens and 15.00 dollars per million output tokens, billed through the Moonshot API. Context cache drops repeated reads to 0.30 dollars per million. That makes K3 roughly 40 percent cheaper than Claude Opus 4.8 and 70 percent cheaper than Claude Fable 5 on list price.

## Kimi K3 Plan Pricing at a Glance

The table below shows the published per million token rates for Kimi K3 next to the Claude models operators most often compare it against, per [eesel AI's Kimi K3 pricing breakdown](https://www.eesel.ai/blog/kimi-k3-pricing) and [Finout's Fable 5 pricing analysis](https://www.finout.io/blog/claude-fable-5-mythos-5-pricing-benchmarks).

| Model | Input per 1M tokens | Cached input per 1M | Output per 1M tokens |
| --- | --- | --- | --- |
| Kimi K3 | 3.00 dollars | 0.30 dollars | 15.00 dollars |
| Claude Opus 4.8 | 5.00 dollars | n/a | 25.00 dollars |
| Claude Fable 5 | 10.00 dollars | n/a | 50.00 dollars |

Access to K3 requires a Moonshot API key from [platform.moonshot.ai](https://platform.moonshot.ai). There is no plan tier to choose from; you pay the metered rates above against your account balance. Kimi.com offers a free consumer chat interface as a separate product from the API tier that charges per token.

The context cache rate of 0.30 dollars per million is the most operationally significant number in that table. Agentic workflows reuse the same system prompt, the same CRM schema, and the same account dossier across multiple steps in a single run. When those tokens hit the cache, input costs drop by 90 percent on the cached portion. [Kimi's K3 documentation](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) confirms the built-in 1 million token context window that makes large dossier stuffing practical without truncation.

## How Kimi K3 Cost Compares to Claude and Codex

The raw rate comparison is straightforward. K3 prices at 40 percent under Opus 4.8 and 70 percent under Fable 5 on list price, per [Morph LLM's K3 versus Claude comparison](https://www.morphllm.com/kimi-k3-vs-claude). That gap holds on both input and output, which matters because agentic work generates a lot of output tokens: planning steps, drafted messages, and structured JSON all count.

Codex sits outside this table. OpenAI does not publish a standalone per token rate for the models behind Codex; pricing runs through ChatGPT plans, making a clean per task comparison harder to construct. If you run heavy volume work and need predictable token costs, K3's published rate is the main reason it enters the conversation. The broader model versus harness question is covered in [Kimi K3 vs Claude Code vs Codex](/blog/kimi-k3-vs-claude-code-vs-codex/).

Two caveats worth stating plainly. First, list price is not the same as cost per completed task. A model that requires more tokens to reach the same output quality can cost more per task even at a lower rate. Second, K3 thinks by default, which adds reasoning tokens that Opus 4.8 would not spend on a call without extended thinking enabled. Measure cost per task on your own workload, not cost per token on a pricing page.

## Cost Per Token in Practice

Numbers are easiest to use when they are already turned into work. The table below shows common agentic tasks with estimated token counts and the resulting cost on K3 versus Claude Opus 4.8. These are worked examples derived from the published rates above. Real numbers move with prompt design and caching. The estimates track the SDR cost analysis in [AI SDR cost on Kimi K3](/blog/ai-sdr-cost-on-kimi-k3/), where the same methodology is applied to a full month of outbound work.

| Task (one unit) | Est. tokens (in / out) | K3 cost | Claude Opus 4.8 cost |
| --- | --- | --- | --- |
| Research one account | 6,000 / 1,000 | ~0.033 dollars | ~0.060 dollars |
| Score and qualify | 2,000 / 300 | ~0.011 dollars | ~0.018 dollars |
| Draft one outbound touch | 3,000 / 350 | ~0.014 dollars | ~0.024 dollars |
| Classify one reply | 500 / 100 | ~0.003 dollars | ~0.005 dollars |

A month of AI outbound work at small team volume, roughly 1,000 accounts researched and 2,000 touches drafted, runs an estimated 55 to 75 dollars on K3. The same workload costs about 100 to 135 dollars on Opus 4.8 and about 250 dollars on Claude Fable 5, based on the per million token rates above. Even if your actual numbers run twice the estimates, K3 still lands below any platform seat in the AI SDR category, where the market spans 600 to 5,000 dollars a month.

Context caching matters at these volumes. If half your input tokens hit the cache on a given agentic run, the 15 million input tokens in that example drop from 45 dollars to about 25 dollars. That outcome is realistic when system prompts, CRM schemas, and company dossiers are reused across research, scoring, and drafting steps within the same run.

## What Cheaper Tokens Mean for Operators

A model price matters most when you can actually pass the savings through to your workflow. Platform seats bundle model cost, orchestration, and margin into one opaque price anchored to the human SDR salary they replace. A platform tied to one fixed model vendor cannot pass a K3 sized price drop through quickly. Its pricing is locked to annual contracts, so savings accrue to the platform margin first and to you at renewal.

A model agnostic runtime inverts this. The model is a swappable config value. When K3 prices at 3.00 dollars per million input tokens and you point your runtime at the Moonshot endpoint, the savings show up on your next bill rather than your next contract renewal. [Yalc's GTM agents for cold email](/gtm-ai-agents/for-cold-email/) follow this architecture: the agent workflow, the CRM context, and the approval rules stay fixed while the model underneath is the layer you optimize for cost.

For the full index of Kimi K3 resources across setup, outbound, and comparisons, see the [Kimi K3 complete guide](/blog/kimi-k3/).

## FAQ

### How much does Kimi K3 cost per million tokens?

Kimi K3 costs 3.00 dollars per million input tokens, 0.30 dollars per million cached input tokens, and 15.00 dollars per million output tokens, per [eesel AI's pricing breakdown](https://www.eesel.ai/blog/kimi-k3-pricing). These rates apply to API calls billed against your Moonshot account balance. There is no separate free tier at the API level; Kimi.com's chat interface is a distinct product.

### Is Kimi K3 cheaper than Claude?

Yes, on list price. K3 runs at roughly 40 percent under Claude Opus 4.8 and 70 percent under Claude Fable 5 on both input and output rates, per [Morph LLM's comparison](https://www.morphllm.com/kimi-k3-vs-claude). Whether it is cheaper per completed task depends on how many tokens each model needs to finish the same job, since K3 thinks by default and that reasoning adds output tokens.

### Does Kimi K3 have a free plan?

Kimi.com offers a free consumer chat product, but the API tier has no free plan. You fund a Moonshot account balance and pay the metered rates per API call. You can start with a small balance to run your own task estimates before committing to volume.

### How do I get access to the Kimi K3 API?

Create an account at [platform.moonshot.ai](https://platform.moonshot.ai), generate an API key, and fund your balance. The base URL is `https://api.moonshot.ai/v1` for standard calls, or `https://api.moonshot.ai/anthropic` if you want to run Claude Code on K3 through the Anthropic compatible endpoint.

### What is Kimi K3 context caching and how much does it save?

Context caching stores repeated portions of your input, typically system prompts and account dossiers, so subsequent calls do not re-bill those tokens at the full rate. K3 caches those reads at 0.30 dollars per million tokens instead of 3.00 dollars, a 90 percent cost reduction on the cached portion. Agentic workflows that reuse CRM schemas or prospect dossiers across multiple steps benefit the most.