Kimi K3 Pricing: API Cost, App Tiers, and What You Actually Pay

Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens through the Moonshot API. Cached input drops to $0.30, a 90% discount. The price is flat across the full one million token context window, with no long-context surcharge.

Those are the numbers. The rest of this article explains what they mean in practice, because the sticker price is only part of the story.

Kimi K3 API pricing

All prices are USD per one million tokens, from Moonshot's official pricing page.

Pricing as of July 2026. Source: Moonshot AI.

Three things about this pricing matter more than the headline numbers.

The cache discount is the real story. A cached input token costs one-tenth of a fresh one. For any workload where you send the same system prompt, documentation, or repository context across requests (coding agents, RAG pipelines, long chats), most input hits the cache. OpenRouter reports a 92% cache hit rate on K3 traffic. At that rate, the weighted average input price drops from $3.00 to about $0.52 per million tokens. Moonshot credits this to its Mooncake split-inference architecture, which handles caching automatically with no separate write fee.

There is one K3 model, with three effort settings. Moonshot does not ship a separate cheaper variant the way OpenAI ships Terra and Luna alongside Sol. What K3 does offer is an effort dial: reasoning_effort accepts low, high, and max, with max as the default. Thinking cannot be switched off entirely at any setting, and reasoning tokens always bill as output at $15 per million. The practical consequence for your bill is that the cheaper settings are opt-in. Leave the parameter unset and every routine query pays maximum-effort reasoning.

Flat pricing across the full 1M context. K3 charges the same rate whether your prompt is 1,000 tokens or 1,000,000 tokens. GPT-5.6 Sol applies a 2x input and 1.5x output surcharge above 272K tokens. Claude and most other current models also price flat, but Sol's surcharge makes K3's flat rate a specific advantage for long-context workloads against OpenAI's flagship.

For a comparison of how K3 competes with GLM 5.2 on price (GLM is roughly 3x cheaper per output token), see our dedicated breakdown.

What K3 pricing means for real workloads

The per-token price is easy to compare. The per-task price is what actually matters for budgeting.

Artificial Analysis measures cost per completed task across its Intelligence Index evaluation suite, which tests coding, reasoning, agentic work, and knowledge retrieval. K3 costs $0.94 per task. For the full Intelligence Index evaluation (all nine benchmarks), K3 cost $2,710 in total and generated 130 million output tokens.

That output volume is the hidden cost driver. K3 is verbose. It generated roughly 2x the median output volume among comparable reasoning models (130 million tokens vs a 63 million median). Cheaper per token, but it uses more tokens. The sticker savings are real but smaller than the rate card implies.

Cost per task: K3 vs the field

All figures from Artificial Analysis, July 2026. Cost per task uses weighted average across nine Intelligence Index evaluations.

K3 sits in the middle of the field on cost per task. It is cheaper than every model that scores higher (Sol, Fable 5) and more expensive than every model that scores lower (GLM, Terra). That is roughly what you would expect from a well-priced model: you pay for what you get.

The comparison that surprises: K3 generates 24K output tokens per task while Sol generates only 15K, yet K3 still costs less ($0.94 vs $1.04). K3 is more verbose but its lower per-token rate ($15 vs $30 output) more than compensates.

What $100 of API spend buys you

Calculated from official output pricing and Artificial Analysis cost-per-task data.

For a team running 1,000 agent tasks per month, the monthly API bill looks roughly like: K3 at $940, Sol at $1,040, GLM 5.2 at $470, Opus 4.8 at $1,800. Those numbers shift with cache hit rates and workload mix, but the ranking is stable.

The always-on reasoning cost

This is the pricing detail most articles skip, and it is the one most likely to surprise your budget.

K3 always thinks. Every request runs a reasoning trace before producing the visible answer, and those reasoning tokens bill as output at $15 per million. You can now change how hard it thinks. You cannot make it stop.

At launch, reasoning_effort accepted only max. It now accepts low, high, and max, with max still the default. For budgeting purposes the important word is default: the lever exists but it is opt-in, so unless you set the parameter explicitly on every call, routine requests pay maximum-effort reasoning.

What that costs in practice: a simple question a non-reasoning model would answer in 200 tokens might cost K3 1,000-3,000 tokens at max effort (200 answer plus 800-2,800 reasoning). That is 5-15x more output tokens for the same visible response. Moonshot has not published token-consumption figures for the low and high settings, so how much you actually save by stepping down is not yet documented.

TrilogyAI's StackPerf benchmark exposed the variance directly, and its findings describe max-effort behaviour specifically. Across 30 identical prompts sent through Kimi Code, reasoning token usage varied by 275%. Four of those requests consumed 4,093 of the available 4,096 completion tokens on reasoning alone and returned no visible answer. The reasoning trace ate the entire budget. Whether the lighter settings reduce that variance has not been tested publicly.

For teams running mixed workloads, Claude and GPT-5.6 retain an edge here, though a narrower one than at K3's launch. Claude Opus 4.8 offers five effort levels (low, medium, high, xhigh, and max) and defaults to high rather than max, so its out-of-the-box cost profile is lower. GPT-5.6 Sol, Terra, and Luna offer effort control from none through max, which includes genuine non-reasoning operation, plus a three-tier family for routing by difficulty. K3 gives you three settings on one model and always thinks.

Kimi K3 compared to other models on price

Sorted by output price, lowest first.

Pricing as of July 2026. Sources: official vendor pricing pages. Claude rates verified against Anthropic's pricing documentation. DeepSeek V4 figures verified against DeepSeek API docs; V4 Pro reflects the standing 75% promotional discount.

K3 lands at the same standard price point as Claude Sonnet 5 ($3/$15), though Sonnet 5 is running introductory pricing of $2/$10 through August 31, 2026, which makes it meaningfully cheaper today. K3 is also a notable step up from its own predecessor: Kimi K2.6 ran roughly $0.95/$4, so K3 is approximately 3.2x more expensive on input and 3.75x on output.

The pricing positions K3 as a capability-priced model, not a budget option. Moonshot is charging for the quality, not undercutting Western models on price. Compared to the K2 generation, the narrative has shifted from "cheap Chinese AI" to "competitive with the Western frontier on both capability and price."

Kimi app subscription tiers

If you are using the Kimi consumer app (not the API), pricing works through subscription tiers. All plans include K3 access.

Pricing from kimi.com/membership/pricing, July 2026.

Two details worth noting. First, Kimi's pricing page carries a banner that new plans are coming and that Kimi and Kimi Code benefits will be separated into different products. If you are subscribing primarily for the coding tool, expect that to become its own subscription. Second, only the Allegro ($99) and Vivace ($199) tiers include K3's full 1M context window. The lower tiers cap context at a shorter length.

For coding-heavy users, the Allegro tier at $99 has drawn favorable comparisons. One Hacker News user reported that the Allegro tier's Kimi Code quota roughly matched what they got from Anthropic's $200 plan when primarily using Fable 5. (Note: Kimi's pricing page indicates Kimi Code benefits may be separated into their own subscription in the future, which would change this comparison.)

The price of self-hosting K3

K3's open weights shipped on July 27, 2026, so the API is now optional. "Free weights" still does not mean "free to run," and in K3's case there are two separate bills to model.

The hardware bill. At 2.8 trillion parameters, K3 is a large model to serve. Moonshot's own guidance recommends 64 or more accelerators. Post-release, SGLang publishes per-hardware serving recipes that start considerably lower: 8 GPUs on B300, GB300, or MI350-class silicon, 16 on B200, GB200, or H200, and 32 on H100. Some of those recipes are still marked unverified. Even at MXFP4, the Hugging Face repository runs to roughly 1.56 TB (1.42 TiB) before KV cache. That is a different class of deployment from GLM 5.2, which fits on a single 8x H200 node at FP8.

The license bill. K3 did not ship under MIT. Moonshot wrote its own document, the Kimi K3 License, tagged license:other on Hugging Face. Internal use is free and unconditional. Two thresholds are not:

• If you give third parties inference or fine-tuning access with control over inputs and parameters, you need a separate agreement with Moonshot once combined licensee and affiliate revenue passes $20 million over any consecutive 12 months
• Products above 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3" in the interface

Access through Moonshot's own products or certified inference partners is exempt. If you are running self-host economics, the $20 million reseller trigger is the one to model, because it lands on exactly the business shape that self-hosts in order to resell capacity. Artificial Analysis classifies the license as requiring a separate agreement for commercial use rather than as an open release.

For most teams, the API remains the practical path. Self-hosting makes sense for organizations with existing GPU clusters, strict data residency requirements, or volume high enough that per-token API cost exceeds the amortized cost of owned hardware, and whose business model does not trip the reseller clause.

One compliance consideration that Layer3Labs flags: the Moonshot API is China-hosted. For regulated industries (healthcare, finance, government), data-residency rules may require self-hosting or a compliant deployment through a third-party provider rather than sending data to Moonshot's infrastructure directly. Together AI is now listed as an inference provider alongside options like Fireworks and OpenRouter, which widens the set of routes available to regulated teams.

For Kimi K3 alternatives that offer open weights at a lower self-hosting bar, see Emergent's dedicated roundup.

Beyond the pricing comparison

If you are comparing K3 pricing to figure out which model to build your app on, there is a simpler path. Emergent lets you describe an application in plain language and get a production-ready, full-stack product with a real backend, real integrations like Stripe, MongoDB, and Shopify, and code you own. It runs Claude, OpenAI GPT, and Google Gemini under the hood through its Universal LLM Key, so you get frontier model capabilities without managing API keys or comparing token prices.

Skip the pricing spreadsheets. Describe your app and let Emergent handle the rest. Start Building and see how far a prompt gets you.