K3.1 Pro's pricing is the easiest part to account for: the standard tier at $2/$10 matches Claude Sonnet 5.5 and GPT-6.1 Sol, is half of Opus 5.5, and is about a third less than Kimi K3 cost at launch. The hard part is the other side: Highspeed costs double, but its "2–3× faster" claim has no verifiable speed figure. This article does three accounts: the standard tier, Highspeed, and the cache discount.
Kimi API standard: $2 in, $10 out per 1M tokens; $1 minimum top-up; OpenAI- and Anthropic-compatible API formats, optional Zero Data Retention. That rate is among the cheapest flagship tiers: it matches Claude Sonnet 5.5 and GPT-6.1 Sol; Opus 5.5 costs about twice as much; it is about a third less than Kimi K3 cost at launch. AA's cost per index task is $2.11, basically level with Kimi K3's $2.00 — a stronger model at the same per-task cost. OpenRouter listed moonshotai/kimi-k3.1-pro the same day; Kimi.com and the mobile apps are a free chat entry point (no published free API tier).
Highspeed: $4/$20, twice the standard rate; Bedrock serves the same Highspeed variant at the Global CRIS standard price of $4/$20, with cross-Region inference, explicit prompt caching, client-side tool calling, structured outputs and Guardrails. The vendor claims 2–3× faster output, but there is no published tok/s figure to date and no separate Highspeed weights repo (whether it is a standalone checkpoint is unclear). Until speed figures appear, Highspeed's value cannot be judged: if you pay double for "faster," you should at least see a verifiable speed number first. Advice: for speed-sensitive long outputs, measure the standard tier on your own tasks first, then decide whether to upgrade; don't pay double in advance for an unverified "2–3×."
The Kimi API offers a 90% cache discount: repeated prompt content bills at a tenth. For 1M-context, long-horizon agent work, that matters more than the headline rate — once system prompts and long documents are cached, later rounds cost far less on input. Bedrock's Highspeed offers explicit prompt caching (needs active configuration). Two caveats: first, the discount saves input only, not output, and K3.1 Pro's default max reasoning effort produces a lot of output tokens; second, LiveBench's per-task cost is $1.395, $2.057 per successful task — the real bill depends on task success rates and thinking load, not the headline rate times token count.
Chat and light use: start with the free Kimi.com and mobile-app entry point; no need to pay first. Regular API work: the standard $2/$10 tier, with the 90% cache discount used fully, is one of the best flagship values available; the default reasoning effort is max, so for long tasks that need cost control, try a lower effort tier (low / high / max). Speed-sensitive long outputs: measure the standard tier's speed first, and evaluate Highspeed only once a published tok/s figure appears — don't pay double in advance for "2–3×." Enterprise: the Kimi API offers optional Zero Data Retention, enterprise access is off by default and needs an admin to enable it; the Bedrock Highspeed variant fits teams already on AWS that need Guardrails.
Double: the standard tier is $2/$10, Highspeed is $4/$20 per 1M tokens (input/output). Bedrock's Highspeed is also $4/$20.
It cannot be verified. The vendor claims 2–3× faster output, but there is no published tok/s figure to date and no separate Highspeed weights repo. Measure the standard tier first, then decide whether to upgrade.
Kimi.com and the mobile apps are a free chat entry point; the Kimi API has no published free tier, with a $1 minimum top-up. OpenRouter bills pay-as-you-go, with no free tier stated.
The Kimi API's 90% cache discount bills repeated prompt content at a tenth — significant for long-context, multi-round work; but it saves input only, not output, and the default max reasoning effort keeps the output bill up.
Reference: Moonshot — kimi.com (Kimi API pricing) · Amazon Bedrock — Kimi K3.1 Pro (Highspeed) · Artificial Analysis — $2.11 per index task · LiveBench — $1.395 per task (rewritten and organized by ModelBrief; the checkout page price is authoritative)