Kimi K3 vs Kimi K2.6: which one should you actually use?

What is the headline difference between Kimi K3 and K2.6?

Kimi K3 is the flagship and K2.6 is the workhorse. K3 has 2.8 trillion parameters, a 1,048,576-token context window and always-on reasoning. K2.6 has 1 trillion parameters and a 262,144-token window. K3 costs $15.00 per million output tokens; K2.6 costs $4.00.

	Kimi K3	Kimi K2.6
Parameters	2.8T	1T
Context window	1,048,576 tokens	262,144 tokens
Input, cache hit	$0.30	$0.16
Input, cache miss	$3.00	$0.95
Output	$15.00	$4.00
Reasoning	Always on, reasoning_effort low/high/max	Thinking and non-thinking modes
Image input	Yes, native	Yes
Video input	Not documented	Yes
Open weights	Yes	Yes
Released	16 July 2026	20 April 2026

Prices are USD per million tokens, excluding tax, from Moonshot’s own pricing pages as verified on 6 August 2026.

Note the one place K2.6 wins on capability: video input. Moonshot documents video support for K2.6 but not for K3, so if you process video, K2.6 is not merely the cheaper option — it may be the only one of the two that works.

Which is cheaper in practice?

K2.6, substantially — but the multiple depends on your input-to-output ratio, which is why a single “3.75× cheaper” figure misleads.

Take a request with 20,000 input tokens and 2,000 output tokens, at a 50% cache hit rate:

• Kimi K3: input costs $0.033, output costs $0.030 → $0.063 per request
• Kimi K2.6: input costs $0.011, output costs $0.008 → $0.019 per request

That is 3.3× cheaper. At 10,000 requests a month, $630 against $191.

Now change the shape. An agent that sends 200,000 tokens of context and gets back 500 tokens shifts the balance toward input pricing, where the gap is 3.2× on cache-miss and only 1.9× on cache-hit. The more your workload is dominated by cached input, the less the model choice matters.

Run your own numbers in the API cost calculator rather than trusting any single ratio.

When do you actually need K3’s context window?

This is the question that decides the choice, and it is easier to answer than it looks. K2.6’s 262,144-token window already holds roughly 196,000 words — about 650 paperback pages, or 26,000 lines of code.

Most workloads never come close. You need K3’s 1,048,576-token window when:

• You are reasoning over an entire codebase rather than selected files.
• You are processing long document sets — legal discovery, full research corpora, complete transcripts — without a retrieval layer.
• You run long agent sessions where the accumulated history plus tool output crosses 262,144 tokens.

If you find yourself chunking and summarising to fit inside K2.6, that is the signal to move to K3. If you are nowhere near the limit, you are paying 3.75× for headroom you never use.

The context window visualizer converts both windows into pages and lines of code if you want to sanity-check against your own data.

Is K3’s reasoning worth paying for?

Sometimes, and it is the harder call. K3 always reasons — you cannot turn it off, only tune it with the top-level reasoning_effort field, which accepts low, high or max and defaults to max.

That default matters for cost. Reasoning tokens are output tokens, billed at $15.00 per million. A K3 request left on max can generate substantially more output than the visible answer suggests. If you move to K3 and your bill is worse than this page’s arithmetic predicted, reasoning_effort is the first thing to check.

K2.6 takes the opposite approach with separate thinking and non-thinking modes, so you can simply not pay for reasoning on requests that do not need it.

Practical rule: if your task has a verifiable right answer that is hard to reach — a subtle bug, a multi-step proof, a genuinely ambiguous specification — K3’s reasoning earns its cost. If your task is extraction, classification, summarisation or formatting, it does not.

What about Kimi K2.7 Code?

If your workload is coding, neither of these is the obvious answer. Kimi K2.7 Code is priced identically to K2.6 on cache-miss input ($0.95) and output ($4.00), has the same 262,144-token window, and is specialised for programming tasks. For coding work it is strictly the better default than K2.6 at the same price.

That makes the real decision tree simpler than a two-way comparison suggests:

1. Coding work → Kimi K2.7 Code
2. Context above 262,144 tokens, or hardest-tier reasoning → Kimi K3
3. Video input → Kimi K2.6
4. Everything else → Kimi K2.6

How do you switch between them?

Both models are served from the same Chat Completions endpoint with the same request format, so in most integrations it is a one-line change:

  response = client.chat.completions.create(
-     model="kimi-k2.6",
+     model="kimi-k3",
      messages=messages,
+     reasoning_effort="low",
  )

Two things to know when moving to K3:

• Set reasoning_effort explicitly. The default is max, which is the most expensive setting.
• K3 adds tool_choice constraints and dynamically loaded tools. These are additive, so existing tool-calling code keeps working, but they are worth adopting if you have a large tool inventory.

Moving in the other direction, from K3 to K2.6, drop reasoning_effort — K2.6 uses thinking and non-thinking modes instead.