Kimi API Pricing: Complete Cost Guide

The Kimi API is billed per token, with no long-context surcharge. Prices range from $0.95 per 1 million input tokens on Kimi K2.6 to $15 per 1 million output tokens on the flagship Kimi K3. Kimi K3 costs $3 per 1 million input tokens and $15 per 1 million output tokens, with a full 1 million token context window.

Our team at Cabina.AI tracks pricing for Moonshot AI directly, since we give developers access to Kimi K3 alongside 55+ other models through one subscription – when a rate changes, we see it firsthand. By our observation, Moonshot retires old Kimi model IDs faster than most Western providers retire theirs, so a rate card that looked current a few months ago can already be pointing at a dead endpoint.

Kimi API Pricing Table – All Models

Moonshot AI, the company behind Kimi, bills the API per token, not per request or per conversation. Kimi api pricing covers four active models as of this writing, in US dollars per 1 million tokens.

That table is the entire current roster. Moonshot AI has retired every earlier model ID: kimi-k2.5 and the whole moonshot-v1 series stopped working on August 31, 2026, the kimi-k2 preview builds were cut on May 25, 2026, kimi-latest was retired in January 2026, and kimi-thinking-preview went first, in November 2025. A retired ID does not throttle or warn – it returns a 404, so a hardcoded model string from an old integration guide can fail silently until someone checks the response.

No Long-Context Surcharge

Kimi K3 charges the same $3 input and $15 output per 1 million tokens whether a request carries 4,000 tokens or fills its entire 1,048,576-token window. That is a real point of difference from Gemini and OpenAI, both of which raise the per-token rate on their flagship models once a request crosses a fixed context threshold. How much is a token in kimi terms stays constant here in a way it does not for either competitor.

Kimi also uses its own tokenizer rather than the one behind GPT or Claude models. Chinese-language content, and other non-Latin scripts more broadly, can come out to noticeably fewer tokens under that tokenizer than under a Western model’s, which matters directly for a workload processing non-English text at volume.

Moonshot AI is based in Beijing, and Kimi API requests are processed on infrastructure the company controls rather than through a Western cloud region. Teams with strict data-residency or vendor-jurisdiction requirements should confirm where their requests get processed before sending regulated or sensitive data through the API, the same review that applies to any provider operating outside a team’s home jurisdiction.

Kimi API Cost Calculator

Pick a model and enter your expected monthly token volume below to estimate a monthly bill, with or without the batch discount where it applies.

Kimi API Request Example

The Kimi API mirrors the OpenAI chat completions format, so the request below looks almost identical to an OpenAI call with a different base URL and API key.

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer your-api-key-here" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {
        "role": "user",
        "content": "What should I search for to find the latest developments in renewable energy?"
      }
    ]
  }'

The same request in Python, using the standard OpenAI SDK pointed at the Moonshot base URL:

from openai import OpenAI

client = OpenAI(
    api_key="your-api-key-here",
    base_url="https://api.moonshot.ai/v1",
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {
            "role": "user",
            "content": "What should I search for to find the latest developments in renewable energy?",
        }
    ],
)

print(response.choices[0].message.content)

Model IDs to use in place of kimi-k3: kimi-k2.6 is the general-purpose alternative. kimi-k2.7-code and kimi-k2.7-code-highspeed are the two coding-focused options.

Pricing Tiers for Moonshot AI, Explained

Batch inference (-40%, Kimi K2 only)

Asynchronous batch jobs price Kimi K2.6 and Kimi K2.7 Code at 60 percent of the standard rate, a 40 percent discount, and batch jobs are exempt from the account’s normal concurrency limit on top of that. Kimi K3 is not on the batch price list at all – the discount only reaches the K2 line, so switching a batch workload to the flagship model gives up the discount entirely.

High-speed inference (Kimi K2.7 Code HighSpeed)

Kimi K2.7 Code HighSpeed runs the same coding-tuned model at roughly double the price of standard K2.7 Code – $1.90 against $0.95 on input, $8.00 against $4.00 on output – in exchange for output around 180 tokens per second. It is the premium option inside the K2 line for anything where response latency, not raw capability, is the bottleneck.

Concurrency and rate-limit tiers

Moonshot groups accounts into tiers based on cumulative recharge, from Tier 0 at a $1 minimum up to Tier 5 at $3,000 recharged. Each tier raises how many requests and tokens the account can process per minute, along with how many requests it can run at once; the per-token price in the table above does not change at any tier. Spending more lifts how fast an account can call the API, not what each call costs.

Minimum prepay, no free trial

An account cannot call the Kimi API at all until it recharges at least $1, and that minimum recharge is non-refundable. There is no signup credit the way some competitors offer one. The closest thing to an incentive is a loyalty voucher: spend $5 cumulatively and Moonshot issues a $5 voucher toward future usage.

Kimi Context Caching

Context caching matches the token sequence at the start of a request and reuses it instead of reprocessing it. Writing a new cached prefix on Kimi K3 costs $3.00 per 1 million tokens for a 5-minute TTL, or $6.00 per 1 million tokens for a 1-hour TTL, billed once per prefix. Reading a cache hit afterward costs $0.30 per 1 million tokens on K3, roughly one-tenth of the $3.00 standard input rate, and every hit inside the TTL window resets the timer at no extra charge.

The short answer to how much do kimi tokens cost on a cache-heavy workload is: the 5-minute TTL breaks even after a single reused prefix, and the 1-hour TTL breaks even after two, so a chatbot or agent that replays the same system prompt across many short-lived requests should almost always have caching turned on.

Moonshot AI does not publish a separate enterprise or volume-discount tier. The concurrency tiers above, unlocked by cumulative recharge, are the only publicly documented way to raise limits beyond the default account.

Which Kimi Model Should You Use?

Kimi K2.6 is the right default for general chat and everyday multimodal work – reading text alongside images or short video in the same prompt. Comparing kimi models api pricing side by side, K2.6 and standard K2.7 Code sit at the identical $0.95 input and $4.00 output rate, so the choice between them comes down to task, not cost.

Kimi K2.7 Code and Kimi K2.7 Code HighSpeed are built for coding workflows specifically, with the HighSpeed variant justifying its roughly 2x price only when a tool genuinely needs faster token output, such as an interactive coding assistant waiting on a visible completion.

Kimi K3 costs three times as much as the K2 line on input and is the model to reach for on long-horizon agentic work or native visual understanding at full precision. From our experience running Kimi at Cabina.AI, teams that default to K3 for routine chat are usually paying flagship rates for a job K2.6 handles just as well.

Is the Kimi API Free?

No. Every account must recharge at least $1 before the API answers a single request, and that minimum is non-refundable. This differs from the free consumer chat product at kimi.com, which is free to use on its own – the API is a separate product with separate billing, and it carries no free tier or trial credit of its own.

Kimi Membership vs Kimi API: Which Is Cheaper?

Kimi sells four consumer membership tiers on top of its free chat product: Andante, Moderato, Allegretto, and Allegro, priced in Chinese yuan at ¥49, ¥99, ¥199, and ¥699 a month rather than in US dollars. Each tier raises a monthly quota of agent uses and parallel agent tasks, plus a separate storage allowance – none of it is a token allowance, and Moonshot has not published a fixed USD price for any of the four.

None of the four membership tiers include programmatic API access. A person chatting with Kimi through the website or app can use the free product or a membership tier; anything built to call a model from code has to go through the API and its own $1 minimum, billed per token instead of by monthly quota.

If you would rather not commit to a single provider before you decide, Kimi is also available on Cabina.AI alongside ChatGPT and Claude. The subscription covers 55+ models in total from one account, which is a lower-commitment way to compare outputs before committing to an API integration.

Kimi API vs ChatGPT API Pricing

Kimi api pricing per million tokens undercuts the ChatGPT API by a wide margin at every tier, though GPT-6 Astra still edges out Kimi K3 on raw context window size.

Use the calculator above to price a specific Kimi model against your own token volume. We cover the full breakdown of every OpenAI model in a separate ChatGPT API Pricing guide on this blog. From running both APIs side by side at Cabina.AI, the gap that actually shows up on an invoice is at the cheap end of each lineup – GPT-5-nano still undercuts Kimi K2.6 on raw per-token price, but Kimi holds a flat rate at any context length where OpenAI long-context pricing can push a large request higher.

Kimi API Price and Model History

Kimi-thinking-preview was the first entry in the current naming scheme to retire, cut in November 2025. Kimi-latest followed in January 2026, then the kimi-k2 preview builds in May 2026, then kimi-k2.5 and the entire moonshot-v1 series on August 31, 2026, the same day the current K2.6 and K2.7 Code models took over as the standing lineup. Tracking kimi api costs per model over that stretch shows Moonshot cycling model IDs roughly every two to three months rather than maintaining one flagship for a year at a time.

Demand briefly outran supply after Kimi K3 launched: as of July 20, 2026, all four paid consumer membership tiers showed as sold out, with Moonshot citing GPU capacity and saying it would reopen spots in batches. The API itself was not affected, only the consumer membership queue.

Kimi API Pricing FAQ