A8gent

Kimi API Pricing for Small Business Automation

Long-horizon agent models on a Singapore endpoint, priced well for coding work but training on your data by default.

Researched by A8gent
Contents
  1. Verdict
  2. API pricing
  3. The lineup as it stands, and why K2 is not what you think
  4. What Kimi costs on two real automations
  5. Agent reliability on a long-horizon lineup
  6. Singapore servers, but read the privacy policy
  7. Where Kimi wins for a small business
  8. How to access
  9. Privacy and residency
  10. What we build
  11. FAQ
  12. Price sources
VerdictShould you run automations on Moonshot AI (Kimi)?

Kimi is built for long-horizon agent work and it shows in the lineup: a million-token flagship, a dedicated coding model at $0.95 in and $4.00 out per 1M tokens, and a cheap multimodal tier at $0.60 and $3.00 as of August 2026. The international platform is operated by Moonshot AI Pte. Ltd. with servers in Singapore rather than mainland China, which simplifies the residency conversation considerably. The thing to read before you build is the privacy policy, which states that the content you input and generate is used to optimise its models, and which contains no explicit opt-out from that use. For automation over your own source code that may be an easy trade. For customer records it is the whole decision.

Best for

Coding and engineering automation, long-horizon agent runs, and teams already using Claude-Code-style harnesses who want a cheaper drop-in.

Not for

Any workload where training-on-your-data by default is unacceptable and you are not going to negotiate an enterprise contract.

Moonshot AI (Kimi) API pricing per 1M tokens

Moonshot prices per million tokens with a large automatic cache-hit discount, typically a fifth to a sixth of the miss price, and a batch API that Moonshot documents as costing 60 percent of the standard model price. Read that carefully: it is a 40 percent discount, not 60 percent off, and it is a distinction worth getting right when comparing against the flat 50 percent discounts published by OpenAI, Anthropic and Google. Batch eligibility is documented for kimi-k2.7-code, kimi-k2.6 and kimi-k2.5, all of which share a 262,144-token window; the K3 flagship is not listed among them. Prices exclude tax, which Moonshot calculates at checkout by jurisdiction. Note also that the earlier Kimi K2 models many people know are no longer the current line, and the legacy moonshot-v1 family was scheduled to sunset at the end of August 2026. Prices below are list prices as of August 2026.

Moonshot AI (Kimi) API pricing per 1M tokens as of August 2026
ModelContextInput / 1MOutput / 1M
kimi-k3 (flagship)Cache-hit input $0.30 per 1M; not listed among batch-eligible models1,048,576$3.00$15.00
kimi-k2.7-codeCache-hit input $0.19; batch-eligible262,144$0.95$4.00
kimi-k2.7-code-highspeedCache-hit input $0.38; exactly double the standard variant for faster output262,144$1.90$8.00
kimi-k2.6 (general)Batch-eligible262,144$0.95$4.00
kimi-k2.5 (multimodal, budget)Cache-hit input $0.10; text, image and video input; batch-eligible262,144$0.60$3.00

List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.

The lineup as it stands, and why K2 is not what you think

Kimi K2 is the model most people outside China have heard of, and it is no longer the current line. As of August 2026 the flagship is Kimi K3, a large multimodal mixture-of-experts model with a million-token context and a thinking mode. Below it sit the K2.7 coding models, a general K2.6, and a cheaper K2.5. The legacy moonshot-v1 family was scheduled to sunset at the end of August 2026.

That churn is worth knowing about because it is the single biggest practical risk of building on a fast-moving provider. If you write an integration against a model name, you are signing up to migrate on the vendor's schedule. The mitigation is straightforward and worth doing regardless of provider: put the model name in configuration, keep an evaluation set of your own real cases, and re-run it whenever you change models. That turns a migration from a fire drill into an afternoon.

The shape of the lineup tells you what Moonshot is optimising for. Having a dedicated code model, a high-throughput variant of it, and a separate general tier is a lineup designed by people thinking about agent loops rather than about chat. The coding model at $0.95 in and $4.00 out per million as of August 2026 is priced to be run continuously, which is exactly what background engineering automation needs.

The flagship is the outlier on cost. At $15 per million output tokens it is priced like a US flagship, and a verbose agent loop will feel that. Use K3 where the million-token context or multimodal input is genuinely needed and run the rest on the K2.x tiers.

What Kimi costs on two real automations

We price every provider on this site against the same two workloads. Workload A is an inbox agent: 1,500 runs a month at about 3,000 input tokens each, of which 2,400 are a stable system prompt and tool schema, plus 400 output tokens. Workload B is document extraction: 4,000 documents a month at 6,000 input and 300 output tokens each.

On kimi-k2.5 at $0.60 per 1M on a cache miss and $3.00 output, Workload A costs about $4.50 a month, dropping to roughly $2.70 once the stable prefix is hitting the automatic cache at $0.10 per 1M. Workload B costs about $18.00, or about $10.80 through the batch API at 60 percent of standard price. On kimi-k2.7-code at $0.95 and $4.00 they are about $6.68 and $27.60, or $16.56 batched. On the kimi-k3 flagship at $3.00 and $15.00 they are about $22.50 and $90.00, with no batch route.

The lesson in those numbers is that the flagship is the expensive part of Kimi and the rest is competitive rather than rock-bottom. K3 at $15 per 1M output is priced like a US flagship, and because it is not among the batch-eligible models there is no discount route for it. Reserve it for the work that genuinely needs a million tokens of context or the strongest reasoning, and run everything else on the K2.x line where both caching and batching apply.

The batch wording deserves care because it is easy to get backwards. Moonshot documents batch inference as costing 60 percent of the standard model price. That is a 40 percent saving, which is less than the flat 50 percent OpenAI, Anthropic and Google publish. It is still worth taking, and batch jobs are documented as exempt from real-time concurrency limits, which for a bulk backlog can matter more than the discount itself.

Agent reliability on a long-horizon lineup

Moonshot describes K3 as its flagship for long-horizon coding and end-to-end knowledge work, and documents configurable reasoning effort, automatic context caching, tool calling, JSON mode and structured outputs across the current line. Those are the mechanics that determine whether an unattended automation parses cleanly and stops when it should. Automatic caching in particular is a quiet convenience: there is no cache_control marker to place and no cache key to manage, which removes a category of implementation mistake.

The published independent research is the right lens for a lineup that markets long-horizon capability. LOCA-bench, released in early 2026, grows agent context under controlled conditions while holding the task fixed, and reports that most models including frontier and open-source ones degrade sharply as context grows, with the spread widening at the long end. A million-token window is a ceiling. Design your agent to compact its context between steps rather than to fill the window because the window exists.

AgentCE-Bench, from April 2026, injects tool-call failures at rates from zero to 0.3 and finds even a 10 percent failure rate visibly degrades performance at every horizon and difficulty tested. Long-horizon work is precisely where that compounds, because a single dropped tool call early in a fifty-step run poisons everything downstream. Retry and backoff, idempotency on writes, hard step limits and a human escalation path are the mitigations, and none of them are model-specific.

One documentation gap to plan around: we could not reach a rate-limit table setting out requests per minute or tokens per minute by tier. If you intend to push a large backlog through, either confirm the limits with Moonshot first or use the batch API, which is documented as not subject to real-time concurrency limits.

Singapore servers, but read the privacy policy

Moonshot runs a genuine international platform: the operating entity is Moonshot AI Pte. Ltd. and the policy states servers are located in Singapore, entirely separate from its mainland China platform. For a UK, EU or US business that removes a real obstacle. A transfer to Singapore is a familiar, well-trodden arrangement that your existing data-protection documentation probably already handles.

The harder issue is not where the data sits but what may be done with it. The published privacy policy states that the prompts, audio, images, videos, files and other content you input and generate while using the products help it optimise its models and understand user needs. In ordinary language, API content feeds model improvement by default. The policy offers opt-outs for advertising identifiers and marketing email and a rights-request email address, but we found no explicit opt-out from training on your content.

We want to state that neutrally, because the discussion around Chinese AI providers generates more heat than light. This is not a claim about competence or intent. It is a factual difference in default terms between Moonshot on the one hand and OpenAI, Anthropic, Mistral and Google's paid tier on the other. That difference may be entirely acceptable for automation over your own codebase or public data, and entirely unacceptable for a workflow processing customer health or financial records. The point is to make the choice deliberately with the terms in front of you.

If Kimi's capability and price are attractive but the training default is not, there are two routes. Ask Moonshot in writing for a contractual opt-out, which is realistic only at meaningful volume and is a negotiation rather than a documented setting. Or run the published open weights yourself, though the flagship's size makes that a multi-node exercise rather than something a small business will do casually.

Where Kimi wins for a small business

The clearest win is engineering automation. If you have a codebase, the coding models are priced low enough to run continuously against it: upgrading dependencies overnight, backfilling tests, applying routine refactors, triaging error reports against the source. That work is usually too expensive to justify on a US flagship and too fiddly to justify a person's time, which puts it in exactly the gap a cheap capable model fills. And crucially, your own source code is data whose sensitivity you can assess yourself, so the terms question is much easier to answer than it is for customer records.

The second win is batch work. The K2.x line runs at 60 percent of standard price through the batch API, a 40 percent saving, and batch jobs are documented as exempt from real-time concurrency limits. A surprising share of business automation is genuinely batch work that has been built as real-time out of habit. Processing a backlog of documents, enriching a database, generating a month of summaries - none of that needs an answer in two seconds, and on a large backlog the concurrency exemption is worth more than the discount.

The third is the Anthropic-compatible endpoint. If your team already runs a Claude-shaped agent harness, pointing it at Kimi is a base URL change. That makes a genuine cost experiment cheap to run: keep your existing setup, swap the backend, run your evaluation set, and compare on your own work rather than on someone's benchmark chart.

Where we would not use it is anywhere the training-by-default terms create a problem you cannot answer to a client. For a business that handles other people's personal data, being able to point at a signed agreement saying the vendor does not train on it is worth more than the token savings. Choose accordingly, and if you want help mapping which of your workflows fall on which side of that line, that is exactly what an automation audit is for.

Strengths for agent workloads

  • A lineup genuinely shaped around agent work, including a dedicated coding model and a high-throughput variant of it
  • 1,048,576 tokens of context on the flagship at flat pricing, with no long-prompt surcharge tier
  • Automatic context caching on every current model, with cache hits at a fifth to a sixth of the miss price and no cache-management code to write
  • A batch API at 60 percent of standard price on the K2.x line, with no real-time concurrency limits applied to batch jobs
  • Configurable reasoning effort, tool calling, JSON mode and structured outputs documented across the current line
  • OpenAI-compatible and Anthropic-compatible endpoints, so it drops into existing harnesses with a base URL change
  • The international platform is operated by Moonshot AI Pte. Ltd. with servers in Singapore, not mainland China

Weaknesses

  • The privacy policy states the content you input and generate is used to optimise its models, and sets out no explicit opt-out from that use
  • Retention is stated as a principle, kept as long as necessary to provide the services, rather than a fixed period
  • The batch discount is 40 percent, less than the flat 50 percent published by OpenAI, Anthropic and Google, despite being described in a way that reads as deeper
  • Flagship output at $15 per 1M is expensive for verbose agent loops that generate a lot of tokens, and the flagship is not batch-eligible
  • The K2.x models cap at 262,144 tokens, so the million-token window only comes with the pricier flagship
  • We could not reach a rate-limit table in the documentation, so entry-tier throughput is something to confirm before planning a bulk job
  • The flagship open weights run to well over a terabyte, so self-hosting it is a multi-node exercise, not a single-server one

How to access Moonshot AI (Kimi)

  • The international platform at platform.kimi.ai, with an API base of api.moonshot.ai, operated by a Singapore entity with servers in Singapore. This is the route to use if you are outside China.
  • The separate mainland China platform at platform.moonshot.cn, which is a distinct service with distinct terms. Do not assume the two are interchangeable.
  • OpenAI-compatible and Anthropic-compatible request formats, which means Kimi drops into Claude-Code-style agent harnesses as a cheaper backend without code changes.
  • OpenRouter and similar aggregators for cross-provider testing.
  • Open weights for the flagship are published on Hugging Face under a Kimi-specific licence. It is permissive for ordinary commercial use, but has conditions for large managed-service resellers and very large consumer products, so read it if either describes you.

Data privacy and residency

  • The international service is operated by Moonshot AI Pte. Ltd. with servers located in Singapore, which is a genuinely different residency position from a mainland-China endpoint and removes the international-transfer complexity that a PRC endpoint creates.
  • The published privacy policy states that the prompts, audio, images, videos, files and other content you input and generate while using the products are used to help optimise its models and understand user needs. In practical terms that means training on API content is the default. This is the most important fact on this page for anyone handling customer data.
  • The policy sets out opt-out mechanisms for device advertising identifiers and marketing email, and a rights-request address at api-service@moonshot.ai, but we found no explicit opt-out from model training on your content. If that use is unacceptable for your workload, the honest position is that you would be relying on a negotiated arrangement rather than a documented setting, so get it in writing before you build.
  • Retention is stated as a principle rather than a period: data is kept as long as necessary to provide the services and for legitimate business purposes, with examples such as account information retained while the account is active. There is no published fixed window for API content.
  • None of this is a security allegation. It is a plain description of the published terms, which differ from OpenAI, Anthropic, Mistral and Google's paid tier, all of which state they do not train on API data by default. Decide with the terms in front of you.
  • For sensitive workloads, the workable pattern is to use Kimi for non-sensitive work such as automation over your own codebase or public data, and route anything containing customer personal data to a provider whose default terms you can accept without negotiation.

What we build on Moonshot AI (Kimi)

  • Coding and repository automation: nightly dependency upgrades, test generation, migration work run against your own code
  • Long-horizon research and document synthesis where the million-token context earns its keep
  • Batch document processing on the K2.x line at 60 percent of standard price, for backlogs that can run overnight
  • Drop-in replacement backend for an existing Claude-compatible agent harness, as a cost experiment
  • Internal tooling automation where the data is your own code and configuration rather than customer records

Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.

Moonshot AI (Kimi) FAQ

How much does the Kimi API cost?

As of August 2026, kimi-k3 lists at $3.00 per 1M input and $15.00 output with cache hits at $0.30, kimi-k2.7-code at $0.95 and $4.00 with cache hits at $0.19, and the multimodal kimi-k2.5 at $0.60 and $3.00 with cache hits at $0.10. The batch API charges 60 percent of the standard price, which is a 40 percent saving, on kimi-k2.7-code, kimi-k2.6 and kimi-k2.5. Prices exclude tax.

What would a real Kimi automation cost per month?

A 1,500-run inbox agent at about 3,000 input and 400 output tokens per run costs roughly $4.50 a month on kimi-k2.5, or about $2.70 with the automatic cache working on the system prompt. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $18.00, or about $10.80 through the batch API. The same job on the kimi-k3 flagship is about $90.00, with no batch route available.

Is the Kimi batch discount really 60 percent?

No, and the wording invites the mistake. Moonshot documents batch inference as costing 60 percent of the standard model price, which is a 40 percent discount. OpenAI, Anthropic and Google all publish a flat 50 percent. It is still worth taking, and batch jobs are documented as exempt from real-time concurrency limits, which on a large backlog often matters more than the price difference.

Do I have to manage caching on Kimi myself?

No. Moonshot documents automatic context caching across the current line, so there is no cache marker to place and no cache key to manage. Cache hits run at a fifth to a sixth of the miss price. The usual discipline still applies: keep the stable material at the front of the prompt and anything volatile, like a customer name or timestamp, at the end, or the prefix will never match.

Is Kimi K2 still the current model?

No. As of August 2026 the flagship is Kimi K3, with K2.7 coding models, a general K2.6 and a budget K2.5 below it. The legacy moonshot-v1 family was scheduled to sunset at the end of August 2026. If you are building an integration, keep the model name in configuration so a lineup change is a config edit rather than a rewrite.

Where is Kimi data processed?

The international platform is operated by a Singapore entity with servers in Singapore, which is separate from Moonshot's mainland China platform. For a UK, EU or US business that is a familiar transfer arrangement rather than a complicated one. Confirm which endpoint your integration actually uses, since the two platforms are distinct services.

Does Moonshot train on API data?

The published privacy policy states that the prompts, audio, images, videos, files and other content you input and generate help it optimise its models. That is training on customer content by default. The policy offers opt-outs for advertising identifiers and marketing email and a rights-request email address, but no explicit opt-out from model training. That is a different default from OpenAI, Anthropic, Mistral and Google's paid tier, and it is the key fact to weigh before sending customer data.

How long does Moonshot keep my data?

The policy states data is kept as long as necessary to provide the services and for legitimate business purposes, with examples such as account information retained while the account is active and violation-related data kept until resolved. There is no published fixed retention window for API content, so if you need a number for a compliance document you will have to request one.

Is Kimi good for coding automation?

That is its strongest use case for a small business. The dedicated coding models are priced to run continuously, and the Anthropic-compatible endpoint means an existing Claude-shaped agent harness can point at Kimi with a base URL change. Your own source code is also data whose sensitivity you can assess yourself, which makes the terms question far easier than it is for customer records.

Can I self-host Kimi?

The flagship weights are published on Hugging Face under a Kimi-specific licence that is permissive for ordinary commercial use but adds conditions for large managed-service resellers and very large consumer products. Practically, the model is large enough that running it is a multi-node deployment, not something a small business will do without a specific reason.