Kimi API Pricing (November 2026)
Kimi API prices run from $2 to $15 per million tokens across 2 current models. The cheapest is Kimi K3.1 Pro at $2.00 input and $10.00 output per million tokens. Every rate below was read on Moonshot (Kimi)'s own pricing page, on the date shown in its row.
| Model | Input /M | Cache read /M | Output /M | Batch in /M | Batch out /M | Context | Status | Checked |
|---|---|---|---|---|---|---|---|---|
Kimi K3.1 Pro kimi-k3.1-pro | $2.00 | $0.20−90% | $10.00 | — | — | 1M | Flagship | 2026-10-28 |
Kimi K3 kimi-k3 | $3.00 | $0.30−90% | $15.00 | — | — | 1M | Current | 2026-10-28 |
Kimi K2.7 Code kimi-k2.7-code | $0.95 | $0.19−80% | $4.00 | — | — | 262K | Deprecated · → Kimi K3.1 Pro | 2026-10-28 |
| Kimi K2.6 | $0.95 | $0.16−83% | $4.00 | — | — | 262K | Deprecated · → Kimi K3.1 Pro | 2026-10-28 |
US dollars per million tokens, list rates on Moonshot (Kimi)'s own API. Cache read is the price of a cached input token. Batch is the asynchronous API, where the vendor offers one. The date in each row is when we last read it on the vendor's pricing page; it links to that page.
Kimi API monthly cost estimates
What light use (10M input + 2M output tokens a month), medium use (100M input + 20M output tokens a month), heavy use (1B input + 200M output tokens a month) costs on each current model, at list price with no caching and no batch discount.
| Model | Input / output per 1M | Light · 10M + 2M | Medium · 100M + 20M | Heavy · 1B + 200M |
|---|---|---|---|---|
| Kimi K3.1 Pro | $2.00 / $10.00 | $40.00 | $400 | $4,000 |
| Kimi K3 | $3.00 / $15.00 | $60.00 | $600 | $6,000 |
Prompt caching and the Batch API cut these figures; the next sections show by how much. For your own token counts, use the LLM API cost calculator.
Kimi API price history
Kimi API price changes and releases
Releases and funding from Moonshot's Kimi line. Sourced from Moonshot's GitHub/model cards, TechCrunch, and Wikipedia — verified at publication.
$2 input, $0.20 cache-hit and $10 output, a third below K3 on every rate, plus a Highspeed variant at twice the price; weights promised by November 4moonshot-v1-auto and the V1 vision previews are discontinued; Kimi points them to K3$3/$15, about three times K2.6's input rateRetired Kimi models and their replacements
12 Kimi models no longer answer on Moonshot (Kimi)'s API. The price is the last one Moonshot (Kimi) listed; the replacement is the current end of the vendor's own succession chain.
| Model | Status | Last price in / out per 1M | Use instead |
|---|---|---|---|
| Kimi K2 (0905 Preview) | Retired May 25, 2026 | $0.60 / $2.50 | Kimi K3.1 Pro |
| Kimi K2 (0711 Preview) | Retired May 25, 2026 | $0.60 / $2.50 | Kimi K3.1 Pro |
| Kimi K2.5 | Retired Aug 31, 2026 | $0.60 / $3.00 | Kimi K3.1 Pro |
| Kimi K2 Turbo Preview | Retired May 25, 2026 | $1.15 / $8.00 | Kimi K3.1 Pro |
| Kimi K2 Thinking Turbo | Retired May 25, 2026 | $1.15 / $8.00 | Kimi K3.1 Pro |
| Kimi K2 Thinking | Retired May 25, 2026 | $0.60 / $2.50 | Kimi K3.1 Pro |
| Moonshot V1 (128K) | Retired Aug 31, 2026 | $2.00 / $5.00 | Kimi K3.1 Pro |
| Moonshot V1 128K Vision Preview | Retired Aug 31, 2026 | $2.00 / $5.00 | Kimi K3.1 Pro |
| Moonshot V1 (32K) | Retired Aug 31, 2026 | $1.00 / $3.00 | Kimi K3.1 Pro |
| Moonshot V1 32K Vision Preview | Retired Aug 31, 2026 | $1.00 / $3.00 | Kimi K3.1 Pro |
| Moonshot V1 (8K) | Retired Aug 31, 2026 | $0.20 / $2.00 | Kimi K3.1 Pro |
| Moonshot V1 8K Vision Preview | Retired Aug 31, 2026 | $0.20 / $2.00 | Kimi K3.1 Pro |
platform.kimi.ai — direct API with an OpenAI-compatible endpoint and automatic context caching (cache-hit input is a fraction of the base rate). Billed per token.
OPEN-WEIGHT · SELF-HOSTKimi K3 (2.8T parameters) and the Kimi K2 line are published on Hugging Face as open weights, self-hostable with no per-token fee; Moonshot says Kimi K3.1 Pro's weights follow by November 4, 2026.
CONSUMER · APPThe Kimi assistant (kimi.com) — a free long-context chat app that made its name handling very long documents. For end-user usage; no API on this surface.
DeepSeek
MIT open weights at the lowest cost-per-token and frontier reasoning quality. The closest rival to Kimi K2's open-weight pitch — text-only, though, where K2 is multimodal.
CHINESE PEERAlibaba (Qwen)
Broad open-weight Qwen catalog with many sizes and modalities. Moonshot counters with a single, very large 1T-parameter open MoE and long-context heritage.
CHINESE PEERZhipu (GLM)
GLM family with free tiers and strong agentic models. Comparable China-residency profile; Kimi leads on raw parameter scale and long context.
PRIMARY RIVALOpenAI
Frontier benchmarks + global ecosystem (GPT-5, ChatGPT). Closed weights and several times pricier than Kimi K2, with no self-host option.
FRONTIER LEADERAnthropic
Leads coding + reasoning with Claude. No open weights and far higher prices — Kimi K2 targets the same agentic-coding use cases at a fraction of the cost.
GLM ARCHITECTZhipu (Z.ai / GLM)
The Beijing lab behind the GLM models, also operating internationally as Z.ai. Spun out of Tsinghua University in 2019, Zhipu ships strong coding/agentic models, multiple free tiers, and open weights — and was the first of China's "Six Tigers" to pursue an IPO. It also sits on the US Entity List.
Kimi API pricing FAQ
Practical questions about Kimi K2, open weights, and long context.
Q · 01 Which Kimi model should I start with?
+
$2.00/$10.00), Moonshot's flagship since October 24, 2026, with a 1M context - cheaper than the K3 it replaces at the top ($3.00/$15.00), which stays on sale. Kimi K2.6 and Kimi K2.7 Code ($0.95/$4.00) are still the cheapest Kimi rates, but they close to new users on October 31, 2026 and are planned to retire by the end of December. See the use case picker above. Q · 02 Are Kimi models really open-weight?
+
Q · 03 What is Kimi's long-context heritage?
+
Q · 04 Where is my data stored?
+
Q · 05 What happened to the older K2 preview models?
+
kimi-k2-0711, 0905, Turbo, and Thinking — retired on May 25, 2026. Kimi K2.5 followed on August 31, 2026, together with the whole Moonshot V1 line (8K, 32K, 128K, moonshot-v1-auto and the vision previews). Moonshot now points every retired model to Kimi K3.1 Pro, and the remaining K2 models - K2.6 and K2.7 Code - close to new users on October 31, 2026, with retirement planned by the end of December.Moonshot AI (月之暗面) was founded in March 2023 in Beijing by Yang Zhilin (a former Meta AI and Google Brain researcher) with Tsinghua University schoolmates Zhou Xinyu and Wu Yuxin. Its consumer assistant, Kimi, launched in October 2023 and built its reputation on very long context — handling roughly 200,000 Chinese characters per conversation, well ahead of peers at the time.
Moonshot's technical signature is the Kimi K2 line: a 1-trillion-parameter mixture-of-experts model with ~32B active parameters, trained with the lab's own MuonClip optimizer and released as open weights (Modified MIT) on Hugging Face. K2.5 (January 2026) added native multimodality via a MoonViT vision encoder, and K2.6 (April 2026) shipped variants from quick chat up to large multi-agent swarms. Kimi K3 (July 2026) moved to a 2.8T-parameter open-weight model with a 1M-token context, and Kimi K3.1 Pro (October 24, 2026), a post-trained K3, became the flagship at a third below K3's price.
Funding has been rapid: an Alibaba-led $1B round in February 2024 (at a $2.5B valuation), rising to a $2B raise at a ~$20B valuation in May 2026 led by Meituan's Long-Z, with Tencent, HongShan, China Mobile, and others among its backers — making Moonshot one of China's top-funded LLM startups.
Pricing on the Kimi Platform is competitive, and the API offers automatic context caching that drops cache-hit input to a fraction of the base rate — meaningful for long-context workloads. The catalog keeps consolidating: the dated K2 previews retired on May 25, 2026, K2.5 and the Moonshot V1 line on August 31, and the remaining K2 models close to new users on October 31, 2026, with retirement planned by the end of December.
The trade-offs: Kimi is China-based, so the direct API stores data under Chinese rules — though the open weights let you self-host anywhere. Versus DeepSeek it competes directly on open-weight frontier quality; versus Qwen it offers fewer model sizes but a distinctive 1T-parameter open MoE and long-context pedigree.