Kimi K2.6

Where does Kimi K2.6 still win?

On price for general work, and on breadth of input. It has the cheapest cache-hit input rate of any model currently open to new accounts, at $0.16 per million tokens, and it documents text, image and video input — the last of which Kimi K3 does not.

That combination makes K2.6 the sensible default rather than the budget option. The flagship is more capable in two specific ways, and if your workload does not touch either of them you are paying 3.8× more per output token for nothing.

	Kimi K2.6	Kimi K3
Context window	262,144	1,048,576
Input, cache hit	$0.16	$0.30
Input, cache miss	$0.95	$3.00
Output	$4.00	$15.00
Video input	Documented	Not documented
Structured output	Not documented	Documented
Non-thinking mode	Available	Always reasons

Per million tokens. Both support text, image, tool calling, JSON mode, context caching, web search and open weights. Our full head-to-head, Kimi K3 vs Kimi K2.6, reaches a verdict rather than leaving it at “it depends”.

When should you move up to Kimi K3?

Two triggers, and only two: your prompts genuinely exceed 262,144 tokens, or you need responses that conform to a schema you supply. Everything else K2.6 does at a fraction of the cost.

The context trigger is easy to test — count your tokens and see. The context window visualizer translates the limits into pages and lines of code if the raw numbers are hard to picture.

The structured output trigger is more often misunderstood. K2.6 documents JSON mode, which guarantees syntactically valid JSON. It does not guarantee your schema. If you are currently validating K2.6 responses against a schema and retrying on failure, that retry loop has a cost, and at some volume K3’s structured output becomes cheaper than the retries even at the higher token price. Worth measuring rather than assuming in either direction.

Why does Kimi K2.7 Code cost the same as Kimi K2.6?

Because Moonshot prices them identically on cache-miss input and output — $0.95 and $4.00 per million respectively. The coding model is not a premium product. That has a blunt consequence: for code, Kimi K2.7 Code is the better default at no extra cost.

The one place K2.6 keeps an edge is cached input, where it is $0.16 against K2.7 Code’s $0.19 per million. On an agent workload dominated by a repeated cached prefix, that is the tier that matters most.

The other place K2.6 wins is web search. Moonshot documents it for K2.6 and not for K2.7 Code, so a coding agent that needs to look things up on its own is a K2.6 job — or a tool-calling job, wiring in your own search.

How much does the cache discount actually save?

A cached input token on K2.6 costs 83% less than an uncached one. That is a large enough gap that prompt-prefix discipline is worth designing around rather than treating as an optimisation to get to later.

The mechanism rewards byte-identical prefixes. A system prompt and a block of tool definitions that repeat unchanged across calls will land in cache. The same prefix with a timestamp, a request id or a reshuffled tool order interpolated into it will not, and every call pays $0.95 per million instead of $0.16.

This is the single most common way we see agent bills come in higher than the estimate. The fix is usually one line — move the volatile value out of the prefix and into the user message, where it belongs anyway. Kimi API pricing explained covers what you are actually billed for, cache tiers included.

Is Kimi K2.6 the better choice for web search right now?

Possibly, and for a reason that has nothing to do with K2.6 being better at it. Moonshot documents web_search for K2.6, and also documents it for Kimi K3 — but notes that K3’s implementation is mid-update and does not currently recommend relying on it. That is Moonshot’s own caveat, not ours.

So on the one capability where you might expect the flagship to lead, the vendor is telling you to be careful. Kimi K2.7 Code does not document web search at all, which leaves K2.6 as the current model with search documented and no warning attached to it.

We have not tested either implementation, so we are reporting Moonshot’s position rather than a result of our own. Treat this as a reason to evaluate before committing, not as a verdict. If search grounding is load-bearing in your product, test it on your own queries regardless of which model you choose.

When should you turn thinking mode off?

Whenever the task does not need reasoning — which is more often than people assume. K2.6 documents both thinking and non-thinking modes, and that choice is something Kimi K3 does not offer at all. K3 always reasons; its reasoning_effort field goes down to low but never to off.

That makes K2.6 the right model for a whole class of work that K3 handles expensively. Classification, extraction, short rewrites, routing decisions, format conversion — none of these benefit from a reasoning pass, and on K3 you pay for one on every single request whether it helped or not.

Reasoning tokens are billed as output tokens. So a workload of ten thousand short classification calls pays for ten thousand reasoning passes it never needed. On a model charging $15.00 per million output tokens rather than K2.6’s $4.00, that compounds quickly.

The pattern worth adopting is to split traffic by task rather than picking one model for the whole application. Route the genuinely hard reasoning to a thinking mode, and let the mechanical work run without one. This is the same argument as the HighSpeed tiering on Kimi K2.7 Code — the expensive option is chosen per request, not per project.

What about Kimi K2.5, which is cheaper still?

It is cheaper, and it is not an option for most readers. Kimi K2.5 undercuts K2.6 at $0.10 cache-hit and $3.00 output per million — but it is closed to accounts registered after the Kimi K3 launch, and Moonshot sunsets the model entirely on August 31, 2026.

If you have an existing K2.5 integration, K2.6 is the natural landing place: same context window, same multimodal input, still has web search, and it is not scheduled for retirement. Our migration guide covers the move.

If you do not already have K2.5 access, the price is academic — you cannot buy it.