Model pricing across providers
Kimi K2.5 API pricing across providers
Kimi K2.5 by Moonshot AI, as priced by 4 API providers. Hosts may serve different versions of the model.
Capability (Epoch AI): Capable · 148.8
Prices per 1M tokens, lowest input price first. Observed –. A dash means no single figure is published.
| Provider | Input | Cached input | Output | Context window |
|---|---|---|---|---|
SiliconFlowKimi-K2.5 · moonshotai/Kimi-K2.5 | $0.45 | $0.07 | $2.25 | 262K tokens |
| Amazon BedrockKimi K2.5 | $0.60 | No single price | $3.00 | 256K tokens |
Novita AIKimi K2.5 · moonshotai/kimi-k2.5 · Serverless tier | $0.60 | $0.10 | $3.00 | 262K tokens |
PPIOKimi K2.5 · moonshotai/kimi-k2.5 · Serverless tier | CN¥4.00 | CN¥0.70 | CN¥21.00 | 262K tokens |
Context windows are each provider's own figure and are never converted: 1M, 1048K and 1049K can all mean 1,048,576 tokens.
About capability scores (Epoch AI)
This model: ECI 148.8 (90% range 147.3–150.4) · range spans bands · Epoch AI score for Kimi K2.5, model dated Jan 27, 2026.
Capability scores: Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks' [online resource], epoch.ai/benchmarks. Accessed Nov 18, 2026. Licensed CC BY 4.0; values here are rounded to one decimal place and grouped into bands.
The Epoch Capabilities Index (ECI) combines many benchmarks on a scale where Claude 3.5 Sonnet is 130 and GPT-5 is 150. Bands, on the displayed one-decimal score: Frontier 160+, Strong 150–159.9, Capable 140–149.9, Entry below 140. Each score has a 90% range, listed with each model's rates and sources; a range that crosses a band boundary is flagged there. Models are not ranked: neighbouring scores often overlap.
A benchmark is not your workload. Epoch's scores mostly reflect runs at the highest reasoning effort, which use more output tokens than default settings. Scores are dated and can shift between Epoch releases. A score belongs to the developer's model; a host listing shows the same model's score, but hosts may serve different versions or quantizations.
How these listings were matched
Same developer model, separate services. Hosts may serve different versions, quantizations, context limits or serving tiers of a model. A listing on several providers is not proof of identical weights or behavior. Labels such as "(US)" are each provider's own row names.
This group was last reviewed on Nov 15, 2026. Each listing is linked by evidence, never by name alone: the provider's official model ID, the model page its price row links, or, for a price list without model IDs (Amazon Bedrock), the price list's own model and developer attribution. The build fails if a listing disappears, is renamed or its model ID, page or attribution changes. Unmatched listings are not shown here. Routed or aggregator listings (e.g. OpenRouter) are not grouped: they resell other providers' capacity rather than host the model themselves.
Headline rates follow the directory's shared rule: the rate a buyer meets by default, never chosen by price. No currency conversion is performed; unknown cache prices are not zero. See the methodology.