moonshot-ai API Pricing

Short answer. Moonshot prices Kimi K3 at $3.00/$15.00 — flagship rates, 20x the input price of its own K2.7-Code at $0.95/$4.00 — and sells a Highspeed variant at exactly double the price for roughly 1.7x the throughput. The gap between its flagship and its coding model is the widest intra-provider spread in this database.

The four models

Model	Input	Output	Out/in	Cache	Context
Kimi K3	$3.00	$15.00	5.00	$0.30 (10%)	1M
K2.7-Code	$0.95	$4.00	4.21	$0.19 (20%)	262K
K2.7-Code Highspeed	$1.90	$8.00	4.21	$0.38 (20%)	262K
K2.6	$0.95	$4.00	4.21	$0.16 (16.8%)	262K

Output at 5x input on K3 is the steepest ratio among the flagships here, which makes long-form generation on K3 markedly more expensive than the headline input price suggests.

Highspeed math

The Highspeed variant doubles both input and output for about 1.7x the throughput (the provider quotes ~180 tokens/s against the standard tier). Per token you pay 2x; per second of wall-clock you pay roughly 1.18x. That is a reasonable trade only when latency is the binding constraint — interactive coding assistants, live review tools — and a pure loss for batch generation.

K3 versus K2.7-Code

K3 is the 1M-context flagship; K2.7-Code is a 262K coding specialist at one-third the input price. For repository work inside 262K tokens, the specialist is dramatically cheaper and, on coding benchmarks, competitive. The flagship earns its price on long-context tasks and general reasoning, not on code.

K3.1 Pro is priced in CNY

Kimi K3.1 Pro does not appear in the USD table above because Moonshot publishes its price on the China platform in renminbi: ¥13 per 1M input, ¥68 per 1M output, ¥1.3 cached input, with a highspeed variant at ¥26 / ¥136 for roughly 2–3x throughput.

At an indicative 7.1 CNY/USD that is about $1.83 / $9.58 — roughly a third below Kimi K3’s $3.00 / $15.00, which is consistent with Moonshot’s own framing of the release as a price cut as well as a capability step.

Two cautions. First, mixing this row into a USD comparison without stating the exchange rate produces a fake ranking. Second, the international USD list had not been published when this record was created, so the CNY figure is the only verified number — treat any USD figure you see elsewhere as an estimate.

Our view

Kimi’s pricing is the most tiered in the market, and buyers should exploit it. Few teams need a 1M-context flagship for every call. Routing code and short-context work to K2.7-Code and reserving K3 for genuinely long inputs cuts spend by 60-70% with little quality loss on the routed tasks.

The 20x input spread between K3 and K2.7-Code is the single biggest intra-provider lever in this database. No competitor offers a comparable step down within one family, and most Kimi buyers do not use it.

Highspeed is priced correctly for what it is and wrongly for what most buyers want. 2x price for 1.7x throughput is a latency product; if your workload is throughput-bound rather than latency-bound, it is a 18% cost increase per unit of work.

A 10% cache rate on the flagship is weaker than it looks. Zhipu and MiniMax offer 18-20%; on a cache-heavy agent loop, K3’s caching saves less than its competitors’ at similar list prices.