AI ModelsJuly 16, 2026

Kimi K3’s $18 API workload ends the budget-model story

By AgentRiot Editorial

Kimi K3 now has a public API price: $3 per 1M uncached input tokens and $15 per 1M output tokens. That makes a standard 1M-input plus 1M-output workload $18, not budget-model territory.

Hero graphic for Kimi K3 with 1M-token context.

Kimi K3 Moonshot AI Grok 4.5 GPT-5.6 Sol Claude Fable 5 AI benchmarks AI pricing long context AI coding

Kimi K3 now has a public API price, and it changes the story.

Kimi’s official API platform lists kimi-k3 at $3.00 per 1M uncached input tokens, $0.30 per 1M cached input tokens, and $15.00 per 1M output tokens, with a 1,048,576-token context window. The listed prices exclude applicable taxes.

For a simple comparable workload of one million input tokens plus one million output tokens, before cache discounts, tools, and subscription fees, K3 costs $18.

That is not Kimi’s old budget position. It is frontier pricing.

K3 is still less expensive than GPT-5.6 Sol or Claude Fable 5 on that raw workload. But it is no longer a cheap alternative to the rest of the market, and it is far more expensive than Kimi’s own current coding API model.

K3 is 3.6 times K2.7 Code on a mixed-token workload

Kimi’s current public pricing for kimi-k2.7-code is $0.95 per 1M uncached input tokens and $4.00 per 1M output tokens. The high-speed version is $1.90 input and $8.00 output.

That produces a clean comparison:

Model	Uncached input / 1M	Output / 1M	Cost for 1M input + 1M output	Context
Kimi K2.7 Code	$0.95	$4.00	$4.95	262,144 tokens
Kimi K2.7 Code Highspeed	$1.90	$8.00	$9.90	262,144 tokens
Kimi K3	$3.00	$15.00	$18.00	1,048,576 tokens

K3 is 3.16 times the uncached-input price of K2.7 Code, 3.75 times its output price, and 3.64 times the cost on the mixed-token workload above. It is not quite a fivefold increase on Kimi’s own current public rates, but it is a decisive move away from value-tier pricing.

The cache story is different. K3 lists $0.30 per 1M cached input tokens, while K2.7 Code lists $0.19. That matters for long-running agent sessions with stable system prompts and repository context. It does not erase the much larger jump in uncached input and output pricing.

The new price sits between Grok and the top-priced frontier models

K3’s $18 workload is expensive relative to Grok 4.5, but lower than GPT-5.6 Sol and Claude Fable 5.

Model	Input / 1M	Output / 1M	Cost for 1M input + 1M output
Grok 4.5	$2	$6	$8
Kimi K3	$3	$15	$18
GPT-5.6 Sol	$5	$30	$35
Claude Fable 5	$10	$50	$60

K3 is 2.25 times the cost of Grok 4.5 for that workload. It is about 49% less expensive than Sol and 70% less expensive than Fable.

That makes “frontier-priced” the useful description. K3 is not charging at the very top of this set, but it is not positioning itself as the inexpensive way to get a large context window either. The API price is a performance claim in commercial form: Moonshot is asking buyers to treat K3 as a higher class of model than K2.7 Code.

Price does not prove that K3 matches Claude or OpenAI on hard tasks. It does tell us Moonshot is no longer selling K3 as a bargain SKU.

The benchmark board K3 still has to join

The commercial signal arrived faster than the independent evidence.

K3 does not yet appear in the current public data from Artificial Analysis or Agent Arena, the two comparative surfaces checked for this article after K3’s API pricing went live. That is not a claim that no K3 benchmark exists anywhere. It is a precise statement about the current public leaderboards buyers can inspect.

Artificial Analysis’s current Intelligence Index v4.1 combines nine evaluations, including Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, and other tests. Its current relevant proprietary-model rows are:

Model / setting	Artificial Analysis Intelligence Index	Rank in current proprietary-model set	Context
Claude Fable 5 with Opus 4.8 fallback	59.86	#1 of 188	1M
GPT-5.6 Sol max	58.89	#2 of 188	1M
Grok 4.5 high	53.83	#8 of 188	500k
Kimi K2.6 (previous Moonshot model)	44.22	separate open-weights comparison set	256k
Kimi K3	Not listed	N/A	1,048,576 tokens

K2.6 is not K3. Its score is only a previous Moonshot baseline, not a substitute for a K3 measurement. The table also does not create a universal ranking: the listed reasoning settings and fallback configurations differ. It does show the public performance map K3 is entering.

Agent Arena provides another useful, distinct check. Its current agent leaderboard lists Claude Fable 5 first, GPT-5.6 Sol second, Grok 4.5 thirteenth, and Kimi K2.7 Code twentieth. K3 is not listed there either.

That leaves K3 in a clear but incomplete position: its API price has reached the frontier band; its independently comparable public score has not.

A 1M window is now an API product, not only a membership feature

Kimi Code already documented K3’s large-context access through upper membership tiers. The API page makes the model’s developer offering concrete: 1,048,576 tokens of context, cached-input pricing, tool calling, structured output, and new K3 API capabilities such as tool-choice constraints and dynamically loaded tools.

That gives buyers a real alternative to the membership plan. It also removes the old excuse for avoiding a direct cost comparison. K3 can now be priced against metered peers.

The remaining question is whether the extra context and K3’s performance justify the move from $4.95 for a K2.7 Code mixed-million workload to $18 for K3. Context capacity helps only when the model can select the right evidence, use tools well, recover from mistakes, and retain constraints over a long run. A larger window is not a benchmark score.

What the tariff does and does not say

K3’s tariff supports one commercial inference: Moonshot believes K3 belongs above K2.7 Code in its product hierarchy and is charging accordingly.

It does not establish that K3 beats Claude, GPT-5.6 Sol, or Grok 4.5. It does not establish that a 1M-token window produces better long-context retrieval or stronger coding-agent behavior. And it does not settle value for a workload with heavy cache reuse, where K3’s $0.30 cached-input rate may change the bill materially.

Moonshot can now be judged on two things at once:

The price is public. K3 is an $18 mixed-million-token model, not a budget upgrade from K2.7 Code.
The independent performance record is still missing. K3 needs disclosed benchmark runs and provider-neutral agent evaluation before buyers can decide whether that price buys frontier results or simply frontier positioning.

For now, the practical buying map is straightforward:

Pick Grok 4.5 when raw API cost matters most. Its $8 mixed-million-token workload is less than half K3’s price.
Pick GPT-5.6 Sol or Claude Fable 5 when their current independent performance evidence justifies the higher bill.
Consider K3 when a 1M context window, Kimi’s API features, and an $18 mixed-million-token cost fit the job. Treat performance parity with the leading models as unproven until K3 reaches public comparative leaderboards.

K3’s price is now clear. The performance receipt is still due.