Moonshot AI Kimi

💰 Pricing & Plans

🔧 API & Developer Experience

• •API Style: OpenAI-compatible REST endpoints with streaming support; the platform.kimi.com console accepts standard Chat Completions payloads, so most OpenAI/Anthropic SDKs and wrappers connect with a base-url swap.
• •Context Caching: Automatic context caching on the platform — repeated system prompts, tool definitions, and long file snippets are billed at the discounted cached-input rate (K3 ¥2/M cached vs ¥20/M uncached) with no manual cache keys.
• •Tool Calls & Agents: K3 supports Tool Calls, tool_choice, dynamically-loaded tools, and JSON Schema output; designed for long-running coding agents that interleave many tool invocations across a single 1M-token conversation.
• •Reasoning Control: K3 always reasons with Preserved Thinking; reasoning_effort accepts low, high, and max (default max). There is no no-thinking mode. Sampling parameters temperature/top_p/n/penalties are fixed, so no-thinking latency needs are better served by the K2 budget models.
• •Vision & Modality: K3 has native vision input alongside text, enabling image + code mixed workloads (screenshots, diagrams, UI mocks) in the same 1M-token context window.
• •SDK & Ecosystem: First-party Python/Node SDKs plus OpenAI-compatible client support; Moonshot also publishes the Kimi K2 weights openly, giving a large self-hosted community and tooling ecosystem.
• •Region & Billing: Billed in CNY; egress to mainland China is direct. The ¥15 new-user coupon does not apply to K3 — account recharge is required before production usage.

🎯 Long-Context Agent & Coding (1M-token Kimi K3)

Kimi's defining strength is long-context agentic coding. Kimi K3 packs 2.8 trillion parameters, a 1M-token context window, and native vision into a single model whose automatic context caching makes repository-scale reasoning economically feasible: a 500K-token codebase held in cache costs ¥2/M cached input instead of ¥20/M uncached, so a coding agent can keep an entire monorepo in context and issue hundreds of tool calls over a long-lived session. This is a different operating point from typical 128K-256K models — K3 treats whole-repository and long-document workflows (multi-file refactors, full-codebase audits, multi-step agent loops) as first-class use cases. The trade-offs are always-on reasoning (reasoning_effort max only) and a premium ¥100/M output price. For teams whose bottleneck is context depth and agentic persistence, K3 is currently one of the strongest domestic options.

🌐 China Access & Latency

Moonshot AI is a Beijing-based company, and Kimi is a native-first mainland China offering: the platform.kimi.com API is directly reachable without a proxy, with no GFW circumvention needed and no cross-border egress fees. Domestic API latency from mainland data centers is typically low single-digit to low-teens milliseconds RTT to Moonshot's China-region endpoints, and billing is in CNY, removing the currency and invoicing friction other global providers carry. This makes Kimi one of the easiest options for China-based production teams needing long-context or agentic coding: sign up with a mainland phone number, recharge in CNY, and call a domestic endpoint. The main caveats are the usual China-cloud concerns — redundancy and SLA tie to Moonshot's mainland infrastructure, and the international experience is secondary to the domestic focus.