Moonshot AI
Operational· checked every few minutes

Kimi K2.5 API

The Kimi K2.5 API for low-cost multimodal agents, 68% below Moonshot's list price

Price per 1M tokens68%off
Input
$0.20$0.60
Output
$0.95$3.00
Cached input
$0.035
Get your API keyTry it in the playground
Model IDkimi-k2.5
chat.py
from openai import OpenAI
client = OpenAI(
    base_url="https://api.zurelay.com/v1",
    api_key="YOUR_ZURELAY_KEY",
)
stream = client.chat.completions.create(
    model="kimi-k2.5",
    messages=[{"role": "user", "content": "Explain quicksort in two sentences."}],
    stream=True,
)
for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="")

Base URL https://api.zurelay.com/v1. Works with any OpenAI SDK. Chat API docs

Context window
256K tokens
Reasoning
none · high
Input
Text, Image, Video
Released
Jan 27, 2026
Status
Operational

Pricing

Kimi K2.5 API pricing. A fraction of the list price.

Pay as you go from prepaid credit, with no subscription. Top up from $10. Requests that fail are never billed.

RateZurelayMoonshot AI list
Input
per 1M tokens
$0.20$0.60
Output
per 1M tokens
$0.95$3.00
Cached input
per 1M tokens
$0.035—

One rate at every prompt length, up to the full 256K context window. Streaming, tool calls and structured outputs cost nothing extra.

Every Kimi model and price
Savings calculator

Move the sliders to your monthly usage.

Input tokens per month50M
Output tokens per month10M
Zurelay
$19.50
Moonshot AI list price
$60.00
You save every month$40.50

$486.00 a year

Overview

What is Kimi K2.5?

Kimi K2.5 is Moonshot AI's open-weight, natively multimodal K2 model for coding, vision and agent work, released in January 2026, with a 262K-token context window and image and video input. The Zurelay Kimi K2.5 API serves it through one OpenAI-compatible endpoint at 68% below Moonshot's list price.

Moonshot AI released Kimi K2.5 in January 2026 and published the weights on Hugging Face under a modified MIT license. Moonshot built it from Kimi K2 with continued pretraining on about 15 trillion mixed visual and text tokens, so it reads images and video natively. It is a Mixture-of-Experts model with 1T total parameters and 32B active per token, with a 400M-parameter vision encoder, MoonViT, and it holds 262,144 tokens of context. On Zurelay the model ID is kimi-k2.5, and it costs $0.20 per 1M input tokens and $0.95 per 1M output tokens.

Moonshot's reported results are strongest on vision and documents. On its own tests K2.5 scores 92.3 on OCRBench, 88.8 on OmniDocBench 1.5, 78.5 on MMMU-Pro and 86.6 on VideoMMMU, a test of understanding video. For coding it scores 76.8% on SWE-Bench Verified and 73.0% on SWE-Bench Multilingual, and it reaches 50.2 on Humanity's Last Exam with tools, against 45.5 for GPT-5.2 and 43.2 for Claude Opus 4.5. These are Moonshot's own numbers.

Moonshot trained K2.5 to write code from visual specs, such as a UI design or a recorded workflow, and to split big tasks across sub-agents that work in parallel. It reports that its agent swarms run up to 100 sub-agents over as many as 1,500 tool calls, and cut run time by up to 4.5 times against a single agent. Kimi K2.6, released in April 2026, is its successor.

Thinking is on by default, and the reasoning comes back as reasoning_content. On Zurelay, set reasoning_effort to none for direct answers without thinking; any other level turns it on. K2.5 reads text, images and video, calls tools, follows a JSON schema in response_format and streams. In our tests it wrote about 49 tokens a second and found one fact buried in a 198K-token prompt. You pay $0.20 per 1M input tokens and $0.95 per 1M output tokens against Moonshot's list price of $0.60 and $3.00, with cached input at $0.035.

Strengths

Where Kimi K2.5 shines. And what teams build with it.

01

Reads documents well

Moonshot reports 92.3 on OCRBench and 88.8 on OmniDocBench 1.5, for scans, tables and dense pages.

02

Understands video

86.6 on VideoMMMU in Moonshot's tests. On Zurelay, send clips as video_url parts next to your text.

03

Low price per token

Costs less than Kimi K2.6, for high-volume work that doesn't need K2.6's extra coding strength.

04

Open weights

Moonshot published the weights under a modified MIT license, so prompts you build on Zurelay carry over to a self-hosted copy later.

Use cases

  • Document and screenshot reading

    Pull text, tables and fields from scans, forms, receipts and screenshots, with JSON schema output for the next step.

  • Code from designs

    Turn a UI mockup or a screen recording into front-end code, the work Moonshot trained it for.

  • Video questions

    Summarize clips, find moments and answer questions about what happens on screen.

  • Agents on a budget

    Tool-calling loops and search agents at a lower price per token than K2.6.

Get started

Call Kimi K2.5 in three steps.

No waitlist and no new SDK. If your code already talks to OpenAI, it already talks to Zurelay.

  1. 1

    Create a key

    Sign up, add credit and create an API key. Set a monthly budget or a rate limit per key if you like.

    Get your API key
  2. 2

    Point your SDK at Zurelay

    Change the base URL. Everything else in your code stays the same.

    https://api.zurelay.com/v1
  3. 3

    Use kimi-k2.5

    Send chat completions as usual. Streaming, tool calls and usage reporting work as you expect.

    kimi-k2.5

Works with the tools you already use

Compare

Similar models. All on the same key.

68%off

Kimi K2.6

$0.30 / $1.30 per 1M tokens

Pick Kimi K2.6, its successor, for stronger coding and agent work: 58.6% against 50.7% on SWE-Bench Pro in Moonshot's tests, at a higher price.

Kimi K2.6 API
60%off

Kimi K3.1 Pro

$0.80 / $4.00 per 1M tokens

Pick Kimi K3.1 Pro, Moonshot's newest flagship, for the hardest coding, 3D and financial work and a 1M-token context, at a much higher price.

Kimi K3.1 Pro API
84%off

DeepSeek V4 Pro 0813

$0.23 / $0.62 per 1M tokens

Pick DeepSeek V4 Pro 0813 for text-only coding and agent work with a 1M-token context.

DeepSeek V4 Pro 0813 API

FAQ

Kimi K2.5 API questions.

Anything else? Ask our team and we’ll answer by email.

How much does the Kimi K2.5 API cost on Zurelay?

Kimi K2.5 costs $0.20 per 1M input tokens and $0.95 per 1M output tokens on Zurelay, with cached input at $0.035. Moonshot's list price is $0.60 input and $3.00 output, so you save 68%. Reasoning tokens count as output, and one rate applies at every prompt length.

Is there a cheaper Kimi K2.5 API?

Yes. Zurelay serves kimi-k2.5 at 68% below Moonshot's list price. You top up prepaid credit and pay only for the tokens you use, with no subscription, and failed requests are never billed.

How do I call the Kimi K2.5 API with the OpenAI SDK?

Install the official OpenAI SDK, set the base URL to https://api.zurelay.com/v1 and use your Zurelay API key. Then pass model "kimi-k2.5" to chat.completions.create; Zurelay also accepts moonshotai/kimi-k2.5. Moonshot recommends temperature 1.0 with thinking on and 0.6 with it off, and top_p 0.95.

What is the Kimi K2.5 context window?

Kimi K2.5 holds 262,144 tokens of context, and in our tests it found one fact buried in a 198K-token prompt. It can write long answers. Reasoning and the answer share the output budget you set with max_tokens, so leave headroom on hard prompts.

Can Kimi K2.5 read images and video?

Yes. Send images in image_url parts and video in video_url parts, as https links or base64 data URLs. Images and video count as input tokens. Zurelay doesn't take PDFs or audio for Kimi K2.5: a request that includes one is turned away with a clear error before anything is billed.

Can I turn off Kimi K2.5's thinking?

Yes. Thinking is on by default, and the reasoning comes back in reasoning_content. Set reasoning_effort to none for direct answers without thinking, which come back faster and use fewer output tokens. Any other level turns thinking on.

Does Kimi K2.5 support tool calling and JSON output?

Yes. Pass tools as you would to any OpenAI model; on Zurelay a tool forced by name is called, and tool_choice "none" works. For JSON, send response_format with json_object or json_schema and the answer comes back as JSON.

How is Kimi K2.5 different from Kimi K2.6?

They share the same size, the same 262K context and the same image and video input. K2.6 is the newer model, and Moonshot reports clear gains in coding and agent work: 58.6% against 50.7% on SWE-Bench Pro and 66.7% against 50.8% on Terminal-Bench 2.0. K2.5 costs less per token, so it suits high-volume vision, document and agent work.

Is Kimi K2.5 on Zurelay the real model?

Yes. Requests go to kimi-k2.5 itself, Moonshot's open-weight model. Zurelay doesn't swap in a smaller or substitute model. Only smart routing, which you turn on yourself, can answer with a close alternative when the model is down, and the x-zurelay-model header names the model that answered. Output is sampled, so wording varies from run to run. Zurelay is an independent service and is not affiliated with Moonshot AI.

More models on Zurelay

All models

Stop paying list price.
Start saving today.

Create a key in seconds, point the SDK you already use at Zurelay, and every request costs up to 90% less from the first token.

Get your API keyTalk to an engineer