15 November 2026 · Site Methodology & Corrections · 5 min read
私たちの Kimi K2 pricing page has carried the same snapshot since October: seven hosts compared, Novita cheapest at $0.12 input / $2 output per 1M tokens. This week's audit pass re-checked that snapshot against the source instead of trusting the file, and two things had changed underneath it.
We pulled OpenRouter's own model page for moonshotai/kimi-k2 directly, and cross-checked it against ArtificialAnalysis.ai's independent provider-pricing page for the same model. Both agree: Novita's current rate is $0.57 input / $2.30 output per 1M tokens.
| Our October snapshot | Checked 15 Nov 2026 | 変化 | |
|---|---|---|---|
| 入力/1M | $0.12 | $0.57 | +375% |
| 出力/1M | $2.00 | $2.30 | +15% |
| Combined (1:1 mix) /1M | $2.12 | $2.87 | +35% |
The input-token jump is the one that matters for most real workloads: Kimi K2's selling point is a 256K context window, which means long prompts and large retrieved documents dominate the token count. A model that got 4.75x more expensive specifically on the thing it's built to consume a lot of is not a rounding error.
ArtificialAnalysis.ai flags the original kimi-k2 snapshot as deprecated and benchmarks it on a slower cadence than active models, pointing instead to Kimi K2 0905 as the current version. The price we had on file wasn't just stale — it was for a model Moonshot has already moved past.
Looking for the current snapshot turned up not one but two newer releases we hadn't tracked at all:
| モデル | 入力/1M | 出力/1M | コンテクスト | Where |
|---|---|---|---|---|
| Kimi K2 (deprecated) | $0.57 | $2.30 | 256K | Novita, verified today |
| Kimi K2 Thinking | $0.60 | $2.50 | 262K | OpenRouter, Fireworks, GCP Vertex, AWS Bedrock — same rate on all four |
| Kimi K2.6 | $0.60 (Moonshot direct) | $2.50 (Moonshot direct) | 262K | OpenRouter marketplace ranges $0.39–$0.65 input depending on backend |
Kimi K2.6 in particular is a different scale of model — Moonshot describes it as a 1-trillion-parameter mixture-of-experts release aimed at agentic, multi-step tool use rather than single-shot chat, which is a different buying decision than "cheapest tokens for the old K2." A page that only shows the deprecated model's price per host is missing the actual current choice a buyer has to make.
We've verified and corrected the one number we checked against two independent sources — Novita's current rate. We have not re-verified the other six hosts on that page (DeepInfra, Groq, Moonshot direct, Together, OpenRouter, Fireworks) one at a time yet, so we're not republishing a new "cheapest host" ranking until that's done properly. Same standard we hold every vendor page to: two independent sources before a number goes live, not a guess extrapolated from one data point.
Check any workload against current numbers with the AI API コスト計算ツール, compare context-window tradeoffs on the RAG コスト計算ツール, or browse the full AI & LLM API category for other providers. New to usage-based LLM pricing? Start with the 無料の API コストガイド.
Novita's Kimi K2 price checked directly against OpenRouter's model page and ArtificialAnalysis.ai's provider-pricing page, 15 November 2026. Kimi K2 Thinking and Kimi K2.6 figures from the same two sources plus Moonshot's own API pricing page. Prices change — confirm current rates before budgeting a production workload.