Kimi K3 Pricing: API Costs, Subscription Plans, and Real-World Tests

Kimi K3 is not sold at one simple price. Developers pay separately for cached input, uncached input, and output tokens, while people using Kimi through its website see a different set of free and paid membership plans. Those two routes are easy to mix up, and doing so can make Kimi look either much cheaper or much more expensive than it really is.

The official international API page currently lists Kimi K3 at $0.30 per million cached input tokens, $3 per million uncached input tokens, and $15 per million output tokens, before applicable taxes. The output rate deserves attention: a concise answer may cost very little, but a long reasoning-heavy report can spend more on generation than on reading the prompt.

Token rates still do not answer the question most people actually have: what will Kimi K3 charge to finish a real piece of work? We therefore separated the published rate card from six practical tests covering a long annual report, conflicting sources, meeting notes, a business chart, a multi-file workflow, and current-source synthesis. Across the six first outputs, Kimi produced five Good results and one Mixed result. The test section reports both cost and answer quality, because a cheap response that needs a full manual rewrite is not good value.

If you want to use Kimi K3 without setting up API billing—and your workflow also spans models such as the GPT 5.6 series, Claude Opus 4.8, GPT Image 2, or Veo 3.1—GlobalGPT offers a simpler way to work. Open Kimi K3 directly in GlobalGPT, switch between leading chat, image, and video models in one workspace, and avoid juggling multiple subscriptions for different tasks.

Essayez Kimi K3 dès maintenant

Kimi K3 Pricing at a Glance

There are three different prices to keep straight. The official API is usage-based. Kimi’s consumer membership is subscription-based. A third-party platform can wrap Kimi K3 into its own credits or multi-model subscription. None of those prices can safely stand in for the others.

For a developer, the API is the clearest route because the bill maps to actual usage. For a non-technical user, membership is easier to budget, but the current checkout situation matters: when we checked on July 21, 2026, Kimi said new membership plans were coming soon and every paid plan card displayed Sold out. The prices below are therefore a dated snapshot, not a promise that the same plans will still be purchasable tomorrow.

Kimi K3 API Pricing Explained

Input token pricing

Input is everything Kimi has to read before answering: your prompt, system instructions, pasted notes, document text, previous messages, and tool results sent back into the conversation. The published uncached rate is $3 per million tokens. A one-million-token context window is a capacity limit, not a flat charge; a 20,000-token request is billed for roughly 20,000 input tokens, not for the full window.

This is why document-heavy work can look surprisingly affordable on the input side. Reading 100,000 uncached tokens costs about $0.30 at the published rate. The catch is that the answer still has its own, higher price.

Output token pricing

Kimi K3 output costs $15 per million tokens, five times its uncached input rate. That does not make an ordinary answer expensive by itself: 10,000 output tokens cost $0.15. It does mean that prompts asking for exhaustive explanations, many long tables, or verbose reasoning can dominate the bill.

In practice, the easiest cost control is often not squeezing another few thousand tokens out of the source document. It is telling the model exactly what the deliverable should contain and how long it should be. A five-section decision brief is easier to budget than “analyze everything in detail.”

Cached input pricing

Kimi lists cache-hit input at $0.30 per million tokens, a 90% reduction from uncached input. The quickstart documentation adds two details that are easy to miss. The previous request must contain more than 256 prompt tokens before a later request can hit the prefix cache, and the long prefix needs to remain unchanged. Caching is automatic for ordinary model requests; there is no cache ID to create.

A sensible use case is asking several questions about the same long report. Put the stable report text first and vary the shorter question after it. A poor use case is repeatedly rewriting the entire prefix, which may prevent a cache hit. Most importantly, do not budget at the cached rate until the API usage data actually records cached tokens.

Kimi K3 cost formula

Total cost =
(uncached input tokens / 1,000,000 × $3.00)
+ (cached input tokens / 1,000,000 × $0.30)
+ (output tokens / 1,000,000 × $15.00)

Taxes and separately priced tools are not included in that formula. When a provider does not return a billed dollar amount, we describe the result as an estimate at official Kimi rates rather than pretending it is an account charge.

If you are comparing Kimi with another document-heavy model API, our Guide des prix de l'IA de Claude provides a useful second rate card for input, output, and subscription costs.

Kimi K3 Subscription Plans

Kimi’s consumer membership page is in transition. It currently says that Kimi and Kimi Code benefits will be separated under new plans, while existing subscribers remain unaffected. At the same time, it still displays the outgoing annual membership structure. That makes the page useful as a current snapshot but risky as evergreen pricing copy.

All prices above were observed on July 21, 2026. Kimi uses RMB and GlobalGPT uses US dollars, and the included features are different. If you also pay for other chat products, our Présentation détaillée des formules d'abonnement à ChatGPT can help you compare the cost of another separate subscription before deciding whether a multi-model plan is better value.

The membership benefits focus on consumer and agent workflows: concurrent tasks, priority, research features, document and slide work, Kimi Code credits, projects, and storage. The comparison table does not present the membership fee as an API token balance. If your goal is to integrate Kimi into a product or process thousands of requests, use the API price rather than dividing a consumer subscription by an imagined number of tokens.

Another important limitation is geographic. These are RMB prices shown on Kimi’s official membership page. They should not be relabeled as global dollar plans, and taxes or local availability may differ. Recheck the page immediately before publication because Kimi has already announced that the plan structure is changing.

Is Kimi K3 Free?

Kimi has a free Adagio membership tier, but that does not make the Kimi K3 API free. Adagio was listed at ¥0 when we checked. It included one concurrent agent task and basic product access, but the membership comparison did not grant it K3’s extra-long chat capacity of up to one million tokens. That specific entitlement appeared only under Allegro.

“Free” should therefore be read as a product tier, not unlimited access to every K3 capability. Model availability, agent limits, queue priority, storage, long-context access, and promotional credits can all differ. API promotions are separate again: a launch rebate or voucher can reduce a temporary bill, but it should not be baked into a permanent cost comparison.

Kimi K3 API vs Subscription: Which Should You Choose?

If you mainly chat, research, and create documents inside Kimi, the free product or a future replacement membership plan will be simpler. If you need reproducible prompts, usage logs, structured output, or an application integration, the API is the honest comparison point. Do not choose the API merely because the per-million-token number looks small; setup, monitoring, and failed requests are part of the practical cost.

For people who regularly move between Kimi, GPT, Claude, Gemini, and search tools, a multi-model workspace can be easier than managing separate subscriptions and API accounts. GlobalGPT currently lists Kimi K3 in its model navigation and sells its own annual Basic, Pro, and Unlimited plans. That is a different product route, not a substitute name for Kimi’s official membership or API.

Estimated Kimi K3 Cost for Common Tasks

The examples below apply the official list prices to assumed token counts. They are useful for budgeting, but they are not our measured test bill.

These numbers reveal a useful pattern. Caching makes repeated work on a stable source dramatically cheaper, but output remains unchanged. A cached 100,000-token report plus a 10,000-token answer costs an estimated $0.18, compared with $0.45 for the first uncached pass. On the other hand, asking for 50,000 output tokens adds $0.75 even before any tool or tax charges.

To compare how another leading model prices similar developer workloads, see our Gemini 3.1 Pro API pricing guide.

Our Real-World Kimi K3 Cost and Quality Test

We designed six tasks around work people might actually pay a frontier model to do: analyze an 85-page annual report, reconcile conflicting sources, turn meeting notes into an executive brief, read a business chart, complete a multi-file workflow, and synthesize a dated source packet. Every task used a fresh context and the first valid output. We did not rerun a weak but valid answer in search of a prettier result.

The annual-report and source-synthesis tasks received three runs each as separate stability checks, producing ten valid runs in total. Testing was conducted through the GlobalGPT CLI with the model listed as kimi-k3 on July 21, 2026. The CLI did not expose cached-token usage, so cached tokens are reported as unavailable rather than zero. Latency includes the complete provider and CLI route and should not be read as a laboratory measurement of Kimi’s native API speed.

Test 1: Annual Report Analysis

Kimi’s first answer turned an 85-page Microsoft annual report into a decision-ready review with six correct headline metrics, page-level citations, five ranked risks, and clear separation between reported facts and interpretation. The answer was strong enough to use after editing, but it was also the slowest and most verbose first run. We rated it Bon, avec Acceptable value.

The three-run stability check was more cautionary. All three runs kept the core financial figures consistent, and runs one and three cited the supporting summary metrics to PDF page 31. Run two cited those same metrics to page 30. We therefore rated annual-report stability Mixte: numerical extraction was reliable, but exact page traceability still needs a human click-through.

Test 2: Conflicting-Source Research

Kimi correctly recognized that apparently conflicting speed figures of 62 and 24 tokens per second came from different measurement environments. It separated vendor benchmarks, independent evaluation, blind preference votes, provider-route data, and a practitioner observation instead of averaging them into one misleading ranking. The answer was detailed but methodologically disciplined, earning Bon et Good value.

Test 3: Meeting Notes to Executive Brief

The meeting-record task tested whether Kimi would quietly turn conditions into commitments. It did not. The answer preserved the October 14 launch as conditional, kept pricing unresolved, separated the EU data-residency review from customer-notice legal review, and retained owners and dates. We rated the result Bon et Good value.

Test 4: Business Chart Analysis

This was the only Mixed first output. Kimi correctly described the revenue direction and noticed visual-design risks, but it failed the exact-number requirement. The chart’s Q4 churn value was 4.8%; Kimi estimated approximately 4.6–4.7%, which also made its Q1-to-Q4 change calculation wrong. The answer showed useful visual reasoning, but not enough precision for an executive metric without manual checking.

Test 5: Multi-File Workflow

Kimi read five isolated files, recalculated four mandatory first-year costs, rejected plans that failed hard requirements, selected the compliant recommendation, and returned a report plus a claim-to-source evidence table. The substance was Bon. We marked value Acceptable because this was one of the slower, more expensive first runs and the CLI could only return the two file bodies; it did not verify that Kimi could write those files autonomously.

Test 6: Current-Source Synthesis

Because the CLI did not expose web-search or URL-fetch tools, this test used a frozen English source packet rather than live browsing. Kimi kept vendor claims, independent benchmarks, blind preference rankings, and provider-route observations separate. It reproduced the pricing example correctly and marked facts after the packet’s July 17 cutoff as unresolved. The first output was Bon et Good value, but it is evidence of source synthesis—not autonomous live research.

The three source-packet runs were consistent on API pricing, benchmark boundaries, access facts, and unresolved claims. Later runs were even clearer about what the packet could not prove after July 17. We rated this stability check Bon.

What Affects Your Kimi K3 Cost?

Kimi K3 always has thinking mode enabled and defaults to max reasoning effort. Its quickstart also fixes temperature, top-p, and several other sampling settings. Those controls are not line items on the price sheet, but they shape how the model works and how much it may generate. The sensible response is to manage the task contract rather than trying to force unsupported sampling settings.

• Document size: more source material raises input usage, especially on the first uncached request.
• Answer length: output is the highest-priced token category, so a focused brief can be much cheaper than an exhaustive report.
• Conversation history: carrying irrelevant messages forward makes each later request larger.
• Tool results: search pages, file contents, and tool responses sent back to the model can become additional input.
• Retries: a second valid generation is a second bill. Quality rerolls should be a deliberate policy, not a reflex.
• Cache hits: stable long prefixes can cut repeated input cost, but only when the API actually records a hit.
• Manual correction: the cheapest response is not the cheapest completed task if every number and citation needs rebuilding.

Our preferred metric is therefore cost per usable deliverable, not simply cost per request. That keeps a two-cent wrong answer from looking more efficient than a forty-cent report that can be sent after a quick check.

Where Can You Access Kimi K3?

The official routes are Kimi’s consumer product, Kimi Code, and the Kimi API Platform. They serve different jobs. The consumer product packages models with agent, research, document, and interface features. The API gives developers a model endpoint and usage accounting. Kimi Code is a coding product with its own credits and workflow.

Third-party API providers can make integration easier or provide one bill across several models, but their prices, latency, caching, tool support, and usage metadata can differ. When discussing a third-party run, label the provider route instead of treating it as a controlled measurement of Kimi’s own infrastructure.

GlobalGPT offers another consumer route: one workspace for chat models, research, image and video tools, and other AI workflows. On July 21, its annual pricing page displayed Basic at $5.80 per month, Pro at $10.80, and Unlimited at $25 during a 50% promotion, and Kimi K3 appeared in the model navigation. Exact model entitlements and promotional prices should still be confirmed at checkout.

Explore Kimi K3 in GlobalGPT’s multi-model workspace if you prefer a consumer interface over direct API setup. Use Kimi’s official API when you need native billing, developer controls, or a product integration.

Is Kimi K3 Worth the Price?

On published rates alone, Kimi K3 is easy to test without taking a large financial risk. A 100,000-token report plus a 10,000-token answer is about $0.45 before taxes and tools. That is inexpensive enough for serious document work, provided the answer is accurate.

The tests make the value case clearer. Kimi returned a usable first answer on all six tasks, with five rated Good and none rated Poor. It was especially effective on long-document extraction, source reconciliation, meeting-to-action conversion, cross-file decisions, and evidence-bounded synthesis. The main quality failure was exact chart reading, and the stability check found one incorrect PDF page citation even when the underlying figures stayed right.

Our verdict is that Kimi K3 is worth testing for document-heavy and analytical work when a person will verify citations and numbers before publication. It is less convincing for unattended workflows that depend on exact visual values. Cost also depends heavily on the access route: the ten-run GlobalGPT ledger total was about $6.774, while the same reported tokens translate to about $2.282 at Kimi’s published API rates. Neither figure includes the value of human review, and the official-rate total is not what GlobalGPT charged.

For casual users, start with the free product and see whether its limits fit. For frequent Kimi users, wait for the new membership structure before committing to an outgoing plan. For developers, the API is the most transparent route. For people whose work spans several model families, compare the total workspace cost and feature coverage rather than buying each subscription in isolation.

Compare GlobalGPT plans if a multi-model subscription fits your workflow better than separate accounts.

Questions fréquemment posées

Combien coûte la Kimi K3 ?

The official international API price is $3 per million uncached input tokens, $0.30 per million cached input tokens, and $15 per million output tokens, before applicable taxes. Kimi’s consumer membership uses separate annual plans, so app pricing should not be treated as API credit.

Is Kimi K3 free?

Kimi lists a free Adagio consumer tier at ¥0, but the free plan has lower product entitlements and does not include the K3 extra-long chat capacity shown for Allegro. Direct Kimi K3 API usage is billed separately by token unless an explicit promotion or account credit applies.

How much is a Kimi subscription?

On July 21, 2026, Kimi displayed annual prices of ¥468 for Andante, ¥948 for Moderato, ¥1,908 for Allegretto, and ¥6,708 for Allegro. All four paid cards were marked sold out, and Kimi said new membership plans were coming soon, so recheck before purchasing.

Does a Kimi subscription include API credits?

The membership comparison describes consumer features, agent usage, and Kimi Code credits; it does not present the subscription as an API token balance. Treat membership and API billing as separate unless Kimi explicitly adds API credits to a plan or purchase confirmation.

How much does Kimi K3 cost for a long document?

At official rates, 100,000 uncached input tokens plus 10,000 output tokens cost about $0.45 before tax and tool charges. That is an estimate, not a universal document price. File extraction, output length, cache hits, retries, and answer quality can change the real workflow cost.

How does Kimi K3 context caching reduce cost?

Kimi charges $0.30 per million cache-hit input tokens instead of $3 for uncached input. The previous prompt must exceed 256 tokens, and the reusable long prefix should remain unchanged. Budget at the cached rate only when the usage data confirms a cache hit.

Does Kimi K3 web search cost extra?

Model tokens and external tool charges should be tracked separately. Kimi’s current quickstart also warns that official web search is being updated and is not recommended for near-term production workflows. Check the latest tool-pricing page before budgeting a search-heavy application.

How can I reduce Kimi K3 API costs?

Keep stable long context at the start of repeated requests, specify a concise output format, remove irrelevant conversation history, and validate prompts on smaller inputs first. Also track cost per usable result: repeated cheap answers can cost more than one accurate, well-scoped generation.

How did Kimi K3 perform in real-world tests?

Five of six first outputs were rated Good and one was Mixed. Kimi performed best on long documents, source reconciliation, meeting briefs, multi-file decisions, and bounded research synthesis. It misread an exact chart value, and one of three annual-report runs cited the wrong PDF page, so visual numbers and citation locations still need human verification.

How much did the ten Kimi K3 test runs cost?

The GlobalGPT transaction ledger recorded 6,774 credits, approximately $6.774, for ten valid runs. Applying Kimi’s published API token rates to the reported 252,055 input and 101,754 output tokens gives a separate reference estimate of $2.282475. Cached-token usage was unavailable, and the reference estimate is not the GlobalGPT bill.