Buying Kimi K3 Tokens: Moonshot Direct vs OpenRouter

Price first, since that's why you're here: Kimi K3 costs the same $3/$15 per million tokens whether you buy from Moonshot's platform directly or through OpenRouter — the router passes requests straight to Moonshot as K3's single provider. The real differences are operational: caching fidelity, rate-limit ownership, billing consolidation and failure behavior. Pick based on those, not price.

Where the two channels actually differ

Context caching. This is the big one for K3 economics. First-party APIs expose caching semantics fully; router pass-through sometimes lags on newer features or reports cache hits differently in usage metadata. On a cache-heavy agent loop, the difference between billed-as-cached and billed-as-fresh input is the difference between roughly $655 and $1,091 a month on our coding-agent preset — see the caching deep-dive. Before committing volume through any channel, send fifty identical-prefix requests and audit the usage objects for cached-token accounting.

Rate limits. Direct accounts get limits from Moonshot, raised on request with usage history. Router traffic shares the router's upstream allocation — usually generous, occasionally congested at launch weeks exactly when you most want capacity. If K3 is on your critical path, owning the relationship (and the limit) is worth the extra account.

Billing and multi-model workflows. One router invoice covering K3, GPT, Claude and whatever you test next month is genuinely convenient, and OpenRouter's per-key spend caps are a clean budget guardrail. Teams already running multi-model routing shouldn't create a Moonshot account just on principle.

Failure modes. Routers add a hop: marginal latency and one more place to have an incident. They also add graceful fallback — automatic rerouting to a backup model — which first-party APIs by definition can't offer. During K3's launch week this cut both ways: router users rode out first-party congestion via queuing, while direct users got Moonshot's status page and nothing else. Decide which failure you'd rather explain.

Data handling. Read both privacy policies, not just one. Direct calls put your prompts under Moonshot's API terms alone; routed calls add the router's logging and retention layer in between. For regulated data, that extra processor in the chain is a compliance question before it's a technical one — and it's the reason some teams pay the operational cost of going direct even while running routers everywhere else.

Our recommendation matrix

Your situation	Buy direct	Use a router
Cache-heavy single-model production	✅
Multi-model product or eval harness		✅
Quick K3 evaluation this week		✅
Need raised rate limits, enterprise terms	✅
Spend caps / one invoice across vendors		✅

After the weights drop

The channel question changes on July 27: open weights mean third-party GPU clouds can host K3 and compete on price for the first time. Router listings will then show multiple providers at different rates — the moment that happens, "same sticker everywhere" stops being true and this page gets a real price comparison table with verification dates, as the pricing page already does for first-party rates. Until then, choose your channel on operations, count your tokens with the token calculator, and model the volume math in the cost calculator.