GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone

Kimi K3 — 1M Context, Open-Weight Flagship

Kimi K3 is Moonshot's flagship model for long-horizon coding and end-to-end knowledge work. Kimi K3 reads a 1,048,576-token context window, always reasons, and takes a three-rung reasoning-effort dial that runs from low to max. Native vision covers images and video, the weights are open, and on reAPI Kimi K3 bills per token, with input, output and cache reads each carrying their own rate, from the same balance as every other model.

Highlights
Moonshot flagship · tool calling · JSON output
Best for
Long-context zh / en work, agents
Input
Chat messages (OpenAI-compatible)
Output
Text, streaming optional
Kimi K3

Ask Kimi K3

Start with the prompt in the composer or choose an example.

Enter to send · Shift+Enter for a new line

Pricing

Credit-based — 1 credit = $0.001 USD. Pay only for completed generations.

First test cost
$0.01515 credits

One 1K-token prompt with a 1K-token answer at the cheapest tier.

Testing budget guide
Add credits
$10
≈ 666 tests
$50
≈ 3333 tests
$100
≈ 6666 tests
Input
$2.50
Price · 1M tokens
Output
$12.50
Price · 1M tokens
Cache read
$0.25
Price · 1M tokens

What is the Kimi K3 API?

Kimi K3 is Moonshot AI's open-weight flagship model for long-horizon coding and end-to-end knowledge work, served on reAPI through an OpenAI-compatible chat completions endpoint. It carries a 1,048,576-token context window and can return up to that entire window in a single response — the default output cap is 131,072 tokens — always reasons at low, high or max effort, and accepts native image and video input alongside tool calling and strict JSON Schema output. Input and output tokens are billed separately per million tokens, and context caching kicks in automatically on repeated prefixes.

What you can build with this model

Real-world workflows and production use cases you can build and ship with this model.

Kimi K3 fixing a layout defect from a screenshot of the broken page

Long-horizon coding, with screenshots in the loop

Moonshot positions Kimi K3 for long-horizon coding, and describes it running engineering tasks with minimal human supervision — working through large codebases and coordinating terminal tools. What separates Kimi K3 from a text-only coding model is that vision is native rather than bolted on: it takes screenshots and visual feedback as input, which Moonshot calls out specifically for game development, frontend engineering and CAD. That closes the loop a text-only model cannot close, because the model can look at what it just built and judge whether the layout is actually wrong before it edits again.

Read the API docs
Kimi K3 reading a large document set and producing a structured deliverable

A million tokens of knowledge work with no chunking layer

The Kimi K3 context window is 1,048,576 tokens, and Moonshot's own FAQ says the rate is flat with no tiering by context length — so filling the window costs what any other token costs. A contract set, a research bundle or a year of filings goes to Kimi K3 whole, which removes the retrieval layer whose only job was working around a small context. Context caching then does the repeat work: it is automatic on Kimi K3, with no cache id, no TTL and no extra parameter, so a stable prefix in front of a changing question is the pattern to build around.

Kimi K3 running an agentic loop with tools loaded partway through

Agents that are made to call tools, not talked into it

Two tool-calling controls arrived with this generation. A required tool choice forces at least one call on the turn, which is how you stop an agent from answering from memory when it was supposed to look something up — the K2 models reject that setting, Kimi K3 accepts it. And a tool can be introduced partway through a conversation by putting its full definition in a system message, so a large tool inventory does not have to be declared up front on every request. Structured outputs and a JSON Schema keep whatever Kimi K3 returns parseable at the end of the loop.

Why reAPI for Kimi K3

Flat rate all the way to a million tokens

Moonshot documents no long-context tier for Kimi K3, so a million-token prompt bills at the same per-token rate as a short one — there is no threshold where the price steps up mid-workload. Context caching is automatic on top of that: a repeated prefix bills at the cache read rate rather than the input rate, with no parameter on your side. No subscription, no minimum commitment, pay-as-you-go from your reAPI balance. The exact numbers are in the pricing table on this page.

OpenAI-compatible drop-in

Kimi K3 is served from a standard `/api/v1/chat/completions` endpoint, and Moonshot's own quickstart drives the model through the OpenAI SDK — so the compatibility is the vendor's design, not a shim. Adopting Kimi K3 is a base URL, a key and a model string. No SDK swap, and nothing about Kimi K3 needs a bespoke client.

One key across Kimi, Claude, GPT and Gemini

The same reAPI key that calls Kimi K3 calls the Claude, GPT and Gemini families too. Putting Kimi K3 head to head with a closed frontier model, or failing over when one provider has a bad hour, is a model string rather than a second integration, a second invoice and a second set of credentials.

Kimi K3 vs Claude Opus 5

On reAPI these two land at effectively the same per-token rate, which makes price the least interesting part of the comparison. What differs is shape: Kimi K3 is the open-weight model that always reasons, hands its reasoning text back and takes video input, while Claude Opus 5 is the closed model with a finer effort ladder that keeps its reasoning to itself. Two of these rows are the ones that break code carried over from another provider — thinking cannot be switched off on Kimi K3, and its reasoning text has to be echoed back in multi-turn histories.

Capability
Kimi K3 on reAPI
Claude Opus 5
Weights
Open — Moonshot calls it the first open-source model in the three-trillion-parameter class
Closed — API access only
Context and output limits
1,048,576-token context; output defaults to 131,072 and can be raised to the whole window
1M-token context; 128k maximum output, fixed
Thinking
Always on and cannot be disabled — lower the effort instead
On by default, and can be disabled at effort high or lower
Effort ladder
Three rungs — low, high, max — defaulting to max
Five rungs from low through max, defaulting to high
Reasoning text
Returned alongside the answer, and required back in the next request
Raw reasoning never returned; a readable summary on request
Visual input
Images and video, as base64 or an uploaded file — public URLs rejected
Images and PDF; no video input

Comparison reflects behavior documented by each vendor at the time of writing. Rates are described as relationships rather than amounts; the live numbers for Kimi K3 are in the pricing table above.

Ship Kimi K3 in three steps

  1. step 01

    Create a reAPI API key

    Sign in to reAPI and create a key under API keys. That one key reaches Kimi K3 and every other model on the platform, billed from a single balance.

    Open
  2. step 02

    Point your client at reapi.ai

    Set your base URL to reapi.ai's API and your key to the one you just created. Any client that already speaks OpenAI Chat Completions works unchanged — no SDK swap is needed to reach Kimi K3.

    Open
  3. step 03

    Send `kimi-k3` as the model

    Set the model field to `kimi-k3` and post your messages. Turn on streaming for long outputs, drop any temperature or top_p you were sending, and pick a reasoning effort — Kimi K3 always thinks, and the default rung is the most expensive one.

    Open
docs/api/kimi-k3

API reference

Drop-in code and the full parameter table.

View full reference
curl https://reapi.ai/api/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      { "role": "user", "content": "Hello" }
    ],
    "reasoning_effort": "high",
    "stream": true
  }'

Parameter validation failed

Rejected synchronously with the failing field named. Check enum values, ranges, and URL fields — nothing is charged.

Authentication or balance issue

401 means a missing or invalid Bearer key; 402 means the reserve exceeds your balance. Manage keys and credits in the console.

Content or material rejected

Reference material and outputs pass automated moderation. Rejected tasks fail with a clear error and a full refund.

Task failed or timed out

A task that reaches a failed state is never charged — the reserve refunds automatically. Keep the task id and retry when ready.

Frequently asked questions

Common questions about this model.

Kimi K3 is billed pay-as-you-go per token from your reAPI balance, with separate input, output and cache read rates and no subscription. The rate does not step up with context length, and reasoning tokens count as output. Current numbers are in the pricing table on this page.

Related models

Explore more models in the same category.

View all models
start building

Ready to ship?

Try it in the playground or grab an API key to integrate now.

Get API keyView API docs