Model overview
K3 is bought for its window: 1,048,576 tokens, four times the 262,144 shared by every other model in the range, so a whole repository or a long agent trace can stay resident instead of being split into passes. That reach is billed for. Input runs $3 per million against $0.95 on the K2.7 coding tier, and output $15 against $4, roughly three times and nearly four times respectively. Cached input at $0.30 is a tenth of the cache-miss rate, the steepest caching discount in the line, which keeps a large fixed prefix affordable across many turns. Reasoning, tool calling, multimodal input and context caching are present throughout the range, so the premium buys window size rather than an exclusive capability.
Billing conditions: Standard Kimi API price. Cache-miss input is shown as Input; automatic cache-hit input is shown as Cached input.
- Context window
- 1,048,576 tokens
- Knowledge cutoff
- —
- Price last verified
- First observed
Descriptions, capabilities, and tariffs come from the linked source. First-party prices take priority; OpenRouter fallbacks are labelled.