Quick Answer
Kimi API pricing, per million tokens: the flagship Kimi K3 costs $3 input and $15 output with a $0.30 cache-hit rate and 1M context. Kimi K2.6 and the coding-tuned K2.7 Code both run $0.95/$4.00 on 256K context. Kimi K2.5 and the legacy V1 family were retired August 31, 2026.
Kimi built its reputation as the cheap open-weight option, and on July 16, 2026, it deliberately broke that reputation: Kimi K3 launched at $3 per million input tokens and $15 output, 3.75 times the output rate of its own K2 line, priced into the frontier conversation it was built to join.
Six and a half weeks later came the quieter event that should worry anyone running a 2025 integration: on August 31, Moonshot retired Kimi K2.5 and the entire legacy V1 family on the same day, and calls to those model strings now return 404s. Unmaintained integrations didn’t quietly overpay; they broke. The original kimi-k2 series had already gone on May 25, 2026, so a 2025 integration may have broken three months before that.
Third-party pages still list the retired models as live, which tells you how fast this vendor moves.
Both facts are why we keep this page current. CloudZero is the first AI and cloud cost platform to integrate directly with Anthropic and the company whose product turns every AI bill into allocated cost per team and outcome, we treat rate cards the way finance eventually has to, as living documents with expiration dates.
Below: the current ladder, the cache and batch rules that halve real bills, and the direct-versus-gateway spread that makes “Kimi’s price” is two different numbers depending on where you buy it.
What is the Kimi API?
Kimi is the model family from Moonshot AI, the Beijing lab behind some of the most capable open-weight releases of the past two years, served through an OpenAI-compatible API at api.moonshot.ai, so existing SDKs switch with a base-URL change. International billing runs in USD; mainland China bills separately in RMB at different rates, a split worth knowing when comparing screenshots.
The K2 line ships open weights under a modified MIT license, while K3’s weights ship under Moonshot’s own Kimi K3 License, worth reading before planning a self-hosted K3 deployment. Either way, which gives Kimi the same dual economics as Qwen and DeepSeek: pay Moonshot per token, or download the weights and pay for your own inference instead. K3, the 2.8-trillion-parameter multimodal flagship, is the reasoning-and-agents tier built for long-horizon coding work across its million-token context.
The current ladder, post-retirement, per million tokens:
| Model | Input (cache miss) | Cache-hit input | Output | Context |
|---|---|---|---|---|
| Kimi K3 (flagship) | $3.00 | $0.30 | $15.00 | 1M |
| Kimi K2.6 | $0.95 | $0.16 | $4.00 | 256K |
| Kimi K2.7 Code | $0.95 | $0.19 | $4.00 | 256K |
Two reading notes on the K2 line. K2.6 and K2.7 Code carry identical headline rates, so choosing the coding specialist costs nothing extra. And a K2.7 Code HighSpeed variant runs 1.90/8.00 with a $0.38 cache-hit rate, double the standard K2.7 Code rate for roughly 180 tokens per second of output, and it is published on Moonshot’s page says so.
Kimi K3 runs $3 per million input tokens and $15 output. Cache hits bill at $0.30, a 90% discount and the deepest on Moonshot’s ladder. Writing a prefix into the cache is billed separately, at $3.00 per million tokens on the five-minute tier and $6.00 on the one-hour tier, so cold agent prefixes cost before any hit saves. Launched July 16, 2026, K3 is the 2.8-trillion-parameter multimodal flagship with a 1M-token context, priced into the frontier tier rather than the value tier the K2 line occupies.
For budget planning, that $15 output rate is the number that matters: K3 costs 3.75 times the K2 line per generated token, and batch discounts don’t reach it. The practical rule: test on K2.6 first, pay K3’s rates only where its quality earns them, and keep prompt prefixes stable so the $0.30 cache rate carries agent workloads.
What happened to Kimi K2.5?
Kimi K2.5’s final published rates were $0.60 input and $3.00 output, until Moonshot retired it on August 31, 2026, alongside the entire moonshot-v1 legacy family. Calls to both now return 404 errors. You can’t buy those rates today, whatever a stale comparison table says.
The retirement carries a lesson bigger than one model. The legacy V1 family had spent months priced above newer, better models (more per input token than K2.6 for half the context and no cache discount), so any 2025 integration nobody revisited was paying more for less right up until the model string stopped resolving.
Retirement is how this market collects technical debt. A 404 at least tells you.
Migration guidance is mechanical: K2.5 traffic moves to K2.6 (nearest general model, $0.35 more per million input) or K2.7 Code for coding loads, and anything quality-sensitive trials K3.
How do Kimi’s caching, batch, and web search billing work?
Caching is automatic, with no configuration. Repeated prompt prefixes bill at the cache-hit rate. Agent and RAG workloads with stable system prompts save 80% to 90% on input, which makes prefix stability the thing to engineer for.
The Batch API charges 60% of standard rates (a 40% discount) on the K2 line for asynchronous work; K3 is absent from the batch pricing page as of this writing, so budget K3 at full rate. Web search bills per successful call on standalone REST endpoints: $0.002 for Web Search Basic, $0.003 for Web Search Pro, and $0.002 for URL Fetch, with failed or empty-result calls free. The older built-in $web_search tool bills $0.005 per call plus the result tokens it injects into subsequent input, and Moonshot expects to deprecate it on October 20, 2026, so new integrations should use the REST endpoints.
Account throughput scales with cumulative recharge across six tiers, from $1 (3 requests per minute) to $3,000 (300 requests per minute). A $1 minimum recharge activates the account.
Why is Kimi’s price different on OpenRouter?
Because “Kimi’s price” is really two markets. Direct from Moonshot, K3 lists at $3/$15. On OpenRouter the same model is served by about twenty independent hosts at anything from roughly $1.40 to $4.50 input and $11 to $22.50 output, with Moonshot’s own endpoint sitting mid-table at $3/$15, so “the OpenRouter price” is a range rather than a number. K2.5, retired at Moonshot, still serves there at $0.45/$2.25 because gateway providers host the open weights independently.
That’s the open-weights economy working as designed: the weights are public, so anyone with GPUs can undercut the lab’s own API, sometimes with different quantization, latency, and reliability. The buyer’s translation: the cheapest Kimi tokens usually aren’t Moonshot’s, the most authoritative ones are, and the difference is a quality-and-support decision, not just a price one. Compare routes the way the model market comparison does, per accepted task rather than per sticker.
Kimi vs. DeepSeek: which is cheaper?
Kimi vs. DeepSeek on headline rates goes to DeepSeek, still the floor of the value tier, covered in our DeepSeek pricing guide. Kimi’s counterargument is the K2 line’s price-to-capability on agentic and coding work, the deep automatic-cache discounts, and K3 as a frontier-adjacent option DeepSeek doesn’t currently field at that scale.
Both ship open weights, both bill from China-jurisdiction platforms internationally, and both reprice fast enough that any comparison older than a quarter is archaeology. The tiebreaker is the same as everywhere in the value tier: run both on two weeks of your real workload and divide by completed tasks. Sticker arithmetic against Grok and the Western tiers lives one link away.
What does a real Kimi workload cost?
A worked example. A coding assistant on K2.7 Code handling 200,000 requests a month, averaging 4,000 input tokens (three-quarters hitting cache on a stable prefix) and 800 output. Cache-miss input: 200M tokens at $0.95 = $190. Cache-hit input: 600M at $0.19 = $114. Output: 160M at $4.00 = $640. Total about $944 a month, with output at 68% of the bill.
Output costs four times input on the K2 line and five times on K3, so capping and shaping responses moves the bill more than model choice within the K2 line.
How do teams keep Kimi spend accountable?
Kimi API spend has all the properties finance teams miss: it bills on a Chinese platform’s invoice outside every hyperscaler bill, it moves through recharges rather than monthly statements, the same model’s tokens may also arrive via gateways on a second bill, and the vendor retires models on 404s. A token-priced line item this dynamic needs watching, not remembering.
CloudZero’s coverage pattern applies unchanged: native connections where they exist, AnyCost for everything else, Moonshot recharges and gateway bills included. Everything lands in the AI Hub beside OpenAI, Anthropic, and the clouds, allocated per team and product, with anomaly detection watching hourly so a retirement-driven migration or a cache-hit-rate collapse shows up as a flagged change, not a quarter-end mystery.
Take the self-guided tour and open the AI Hub first. Schedule a demo and ask for that view.
Frequently asked questions about Kimi API pricing
