AI Models

What Is Kimi K3 and How Much Does the API Cost?

Published 2026-07-29 · Researched and price-checked on official sources

Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight model. Telnyx lists it at $2.70 per million input tokens, $0.27 per million cached input tokens, and $13.50 per million output tokens as of July 29, 2026. Its custom license has separate conditions for large Model-as-a-Service operators and very large products.

Key facts

What matters

Telnyx lists Kimi K3 at $2.70/MTok input, $0.27/MTok cached input, and $13.50/MTok output.
Moonshot documents a 2.8T-parameter mixture-of-experts model, 104B activated parameters, native vision, and a 1,048,576-token context window.
The weights are published on Hugging Face under the custom Kimi K3 License, not an OSI standard license.
The $20M twelve-month revenue condition applies specifically to Model-as-a-Service operators; a separate attribution condition applies to products above 100M monthly active users or $20M monthly revenue.

What Is Kimi K3?

Kimi K3 is a multimodal agentic model developed by Moonshot AI. Its official model card describes a mixture-of-experts architecture with 2.8 trillion total parameters and 104 billion activated parameters. It uses Kimi Delta Attention and Attention Residuals, supports text and image input, and lists a 1,048,576-token context window.

Moonshot publishes the full model weights on Hugging Face and provides deployment examples for vLLM, SGLang, and Docker Model Runner. Those instructions make self-hosting possible, but the model's size means buyers should size infrastructure from the official deployment guidance instead of assuming it fits a single workstation.

Kimi K3 API Pricing on Telnyx

Telnyx added Kimi K3 on July 28, 2026. Its release note lists model ID moonshotai/Kimi-K3, an OpenAI-compatible API, prompt caching, tool calling, structured output, and three reasoning-effort settings. Verified pay-as-you-go pricing as of July 29, 2026:

Cached input tokens: $0.27 per million
Standard input tokens: $2.70 per million
Output tokens: $13.50 per million

Telnyx's public pricing table shows one rate set for Kimi K3. Buyers with committed volume should check the provider's current contract pricing separately.

Kimi K3 API Pricing Compared

For a clean comparison, use the vendors' standard short-context list prices and keep provider-specific discounts separate. Telnyx lists Kimi K3 at $2.70 input, $0.27 cached input, and $13.50 output per million tokens. OpenAI lists GPT-5.6-terra at $2.50 input, $0.25 cached input, and $15 output. Anthropic lists Claude Sonnet 4.5 at $3 input, $0.30 cache reads, and $15 output.

Kimi K3 is therefore not the cheapest option in every column. Its standard input price is $0.20 above GPT-5.6-terra, while its output price is $1.50 lower. The comparison is only a billing reference: context rules, cache writes, latency, model quality, and provider features can matter more than a small token-price difference.

Anthropic's separate $2/$10 introductory offer through August 31, 2026 applies to Sonnet 5, not Sonnet 4.5. It should not be used as a Sonnet 4.5 price.

Is Kimi K3 Actually Open Source?

Moonshot calls Kimi K3 an open model and publishes the weights, but the repository uses a custom Kimi K3 License rather than MIT, Apache 2.0, or another OSI standard license. The distinction matters when a company plans to resell access or operate at very large scale.

The key terms:

The license permits use, copying, modification, distribution, fine-tuning, and sale, subject to its stated conditions.
A Model-as-a-Service operator whose aggregate revenue exceeds $20 million over any consecutive 12 months must enter a separate agreement with Moonshot before using the software commercially.
A commercial product with more than 100 million monthly active users or more than $20 million in monthly revenue must display “Kimi K3” prominently in its interface.
The license exempts internal use and use through Moonshot's official products or certified inference partners from those two conditions.

This is a practical summary, not legal advice. Teams near either threshold should review the license text and obtain counsel instead of relying on a short article.

Who Should Use Kimi K3?

Kimi K3 is most relevant to three groups:

API buyers comparing multimodal long-context models. Telnyx offers an OpenAI-compatible endpoint and publishes separate cached-input pricing.
Infrastructure teams capable of serving very large open weights. Moonshot provides official vLLM, SGLang, and Docker deployment examples, but hardware sizing remains the operator's responsibility.
Teams building multi-provider routing. Kimi K3 adds another documented API option with a different price profile and a self-hosting path.

Model-as-a-Service operators near the license revenue condition should resolve licensing before deployment. Other teams should still compare latency, output quality, support, data handling, and the cost of cache writes—not only headline token prices.

At a glance

Model	Input	Cached Input / Read	Output
Kimi K3 (Telnyx)	$2.70	$0.27	$13.50
GPT-5.6-terra (OpenAI)	$2.50	$0.25	$15.00
Claude Sonnet 4.5 (Anthropic)	$3.00	$0.30	$15.00

Standard short-context list prices per 1M tokens, checked July 29, 2026. Cache-write charges and discounts are excluded.

FAQ

Is Kimi K3 free to use?

The weights can be downloaded from Hugging Face, while hosted inference is billed by the provider. Telnyx lists $2.70 per million input tokens, $0.27 per million cached input tokens, and $13.50 per million output tokens. Large Model-as-a-Service operators should review the custom license's revenue condition.

How does Kimi K3 pricing compare with GPT-5.6-terra?

At standard short-context list prices, Kimi K3 on Telnyx costs $2.70 input, $0.27 cached input, and $13.50 output per million tokens. OpenAI lists GPT-5.6-terra at $2.50 input, $0.25 cached input, and $15 output. This compares billing only, not model quality or latency.

Can I run Kimi K3 on my own server?

Yes, if you have infrastructure capable of serving the model. Moonshot publishes the weights and official examples for vLLM, SGLang, and Docker Model Runner. Review the custom license and size hardware from the official deployment guidance before treating self-hosting as the cheaper option.