Cheapest Kimi Model: Base K2 at $0.57/$2.30 per 1M

The cheapest Kimi tier on a per-token basis is the base K2 model (the OpenRouter kimi-k2 alias, generation K2-0711) at $0.57 in and $2.30 out per million tokens, measured via OpenRouter; every richer tier costs more per token, and the reasoning tier can cost more per task. That last clause is the part most buyers miss, and we have the first-hand numbers to show why.

If you are picking a Kimi tier purely to save money, the answer is short: start at the base K2 model and only move up when a task genuinely needs the extra capability. The harder question is what "cheapest" means once a reasoning tier is in the mix, because per-token rank and per-task rank do not always agree. We will rank the tiers both ways below.

Kimi tier pricing ladder (verified 2026-06)

Here is the full ladder, cheapest first. The four Kimi K2 rows are measured via OpenRouter, because the native Moonshot key in our environment only reaches the legacy moonshot-v1 family, not the K2 generation. The flagship K2.6 row is official Moonshot pricing and carries a native-reverify flag.

Tier	Input ($/1M)	Output ($/1M)	Source
Base K2 (`kimi-k2` alias = K2-0711)	$0.57	$2.30	measured via OpenRouter
K2-0905	$0.60	$2.50	measured via OpenRouter
K2-thinking	$0.60	$2.50	measured via OpenRouter
K2.7-code	$0.74	$3.50	measured via OpenRouter
K2.6 (flagship)	$0.95	$4.00	official Moonshot pricing

The base K2 model is the genuine floor. It is the tier the OpenRouter kimi-k2 alias resolves to, and it is the generation tagged K2-0711, not the flagship K2.6. People conflate those two constantly, so it is worth saying plainly: the alias you reach when you type kimi-k2 is the older, cheaper generation, and that is the one ranked cheapest here.

According to Moonshot AI's platform, the current flagship is Kimi K2.6 at $0.95 input and $4.00 output per million tokens with a 262,144-token context window, which is the most expensive row on the ladder and the one we could only source rather than route through our own key. We mark it needs_native_reverify for that reason; treat it as sourced, not first-hand.

One honest caveat: OpenRouter is not cheaper for base K2

A pattern holds for several Chinese models where a third-party route undercuts the official price. Kimi K2 is the exception, and pretending otherwise would mislead you. According to Moonshot AI pricing, the official legacy K2 (0711) rate is $0.55 input and $2.20 output per million tokens, which is slightly below the $0.57 / $2.30 we see routed through OpenRouter. The gap is small, a couple of cents per million, but it runs the opposite direction from the usual story. If you have native K2 access, the official endpoint is marginally cheaper; the OpenRouter route buys you convenience and aggregation, not a discount. We say this because the cluster's job is a neutral verdict, not a referral pitch.

First-hand evidence: the cheapest per-token tier is not always the cheapest per task

Documentation gives you the rate card. It does not tell you how many tokens a tier will actually spend on your prompt, and that is where the reasoning tier surprised us. When we called the endpoints via OpenRouter on the same short, two-sentence general prompt, the per-token ranking inverted at the task level.

Tier (via OpenRouter)	Output tokens	Billed cost	Latency
Base K2 (`kimi-k2`)	60	$0.00015852	7.03s
K2-thinking	553	$0.0014041	18.51s

On paper K2-thinking ($0.60 / $2.50) sits only a few cents above base K2 per token, so you would expect a near-tie. Instead the reasoning tier emitted 553 output tokens to base K2's 60 on the identical prompt, and at 18.51 seconds against 7.03 it was both the slowest and, at $0.0014041, roughly nine times the cost of the base-K2 call. The verbosity, not the per-token rate, is what made it the most expensive single call we logged. This is the practical upshot: reasoning tiers bill by the thinking they do, and a chatty reasoner on a simple task can quietly outspend a flagship.

According to Moonshot AI documentation, K2-thinking is a reasoning variant that produces extended chain-of-thought before its answer, which is the mechanism behind that 553-token output. The lesson is not "avoid the reasoning tier." It is "match the tier to the task," and budget reasoning tiers by expected output volume, not by their headline per-token rate.

Per-token rank versus per-task rank

So which tier is cheapest? It depends on the question.

Per-token, the ranking is fixed: base K2 ($0.57 / $2.30) is cheapest, then K2-0905 and K2-thinking tie at $0.60 / $2.50, then K2.7-code ($0.74 / $3.50), then flagship K2.6 ($0.95 / $4.00). This is the right lens for high-volume, predictable workloads where output length is stable.

Per-task, output verbosity reorders things. A tier that spends more tokens per response can cost more in practice even at a similar rate, which is exactly what we saw with K2-thinking against base K2. For latency-sensitive or simple-prompt workloads, the cheapest real choice is the one that answers concisely, and on our test that was base K2 by a wide margin.

For comparison shoppers crossing vendor lines, DeepSeek's V4-Flash sits well below every Kimi tier on the rate card. According to DeepSeek API documentation, V4-Flash bills $0.14 input and $0.28 output per million tokens, roughly a quarter of base K2's input rate and an eighth of its output rate. If raw price is your only axis, DeepSeek wins; if you need Kimi's 256K-context behavior, you are paying the Kimi ladder. We cover that trade-off in depth in the DeepSeek API pricing hub.

Choose this tier if

Choose base K2 if you want the cheapest Kimi entry point and your prompts are short to medium length. Choose K2.7-code if you are doing code generation that benefits from the code-tuned tier and you accept the $0.74 / $3.50 rate. Reach for K2-thinking only when a task genuinely needs multi-step reasoning, and size your budget against its output volume, not its per-token rate. Avoid the flagship K2.6 unless you specifically need its capability, since it is the priciest row and the only one we could not route first-hand.

FAQ

What is the cheapest Kimi model? The base K2 model, reached through the OpenRouter kimi-k2 alias (generation K2-0711), at $0.57 input and $2.30 output per million tokens measured via OpenRouter. It is the floor of the Kimi ladder; every other tier costs more per token.

Is the kimi-k2 alias the same as flagship Kimi K2.6? No. The kimi-k2 alias resolves to the K2-0711 generation, an older and cheaper tier. The flagship K2.6 is a separate, more expensive model at $0.95 / $4.00 per million according to Moonshot pricing. Do not conflate them when budgeting.

Is OpenRouter cheaper than official Moonshot for K2? No, and we will not pretend it is. Official legacy K2 (0711) is $0.55 / $2.20 per Moonshot pricing, slightly below the $0.57 / $2.30 we measured via OpenRouter. The two are near parity, with the official endpoint marginally cheaper.

Why did K2-thinking cost more than base K2 on the same prompt? Because it is verbose. On our OpenRouter test of an identical two-sentence prompt, K2-thinking emitted 553 output tokens versus base K2's 60, at 18.51 seconds and $0.0014041, roughly nine times the base-K2 cost. Reasoning tiers bill by the thinking they generate.

Which Kimi tier is cheapest for coding? For code work the code-tuned K2.7-code tier ($0.74 / $3.50 via OpenRouter) is purpose-built, but plain base K2 is cheaper per token if the task does not need the code tuning. Test both on your actual prompts.

This is part of the Kimi API pricing hub, where we rank every Kimi tier and compare them against DeepSeek and the major US models. Kimi K2 figures are measured via OpenRouter and the flagship K2.6 row is sourced from Moonshot pending native re-verification.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-06-26.