Kimi’s K2 family has several variants — Thinking, Turbo, and version bumps like K2.5 and K2.6 — and they don’t all cost the same. If you’re using Kimi to keep coding costs down, it pays to understand how Moonshot bills, where the cache discounts kick in, and why the same model can show different prices depending on where you get it.
This explains the structure and gives the official rates so you can estimate your real spend. For setup, see run Kimi K2 Thinking with Claude Code.
Pay-per-token, no flat plan
Kimi is billed per token — input and output, per million — with no flat coding plan in the GLM or Alibaba sense. That makes it well suited to light or variable use: you pay only for what you run, and the rates are low. The flip side is no fixed ceiling, so heavy use scales linearly (minus cache discounts).
Official rate snapshot
These prices were checked on June 9, 2026 from Moonshot/Kimi’s official pricing docs:
Kimi / Moonshot pricing (per 1M tokens)
| kimi-k2.6 | Cache-hit input: $0.16; cache-miss input: $0.95; output: $4.00; context: 262,144 tokens |
|---|---|
| kimi-k2.5 | Cache-hit input: $0.10; cache-miss input: $0.60; output: $3.00; context: 262,144 tokens |
| moonshot-v1-8k | Input: $0.20; output: $2.00; context: 8,192 tokens |
| moonshot-v1-32k | Input: $1.00; output: $3.00; context: 32,768 tokens |
| moonshot-v1-128k | Input: $2.00; output: $5.00; context: 131,072 tokens |
The variants and what drives their cost
Kimi K2 variants
| K2.6 | Latest listed K2 generation; higher cache-miss and output rate than K2.5, stronger agent/code-writing positioning |
|---|---|
| K2.5 | Cheaper K2 generation at $0.60 cache-miss input / $3.00 output per 1M tokens |
| Thinking mode | Cost rises mostly because it generates more reasoning/output tokens; reserve it for hard tasks |
| Turbo / fast variants | Use for routine coding if available from your provider; compare the provider's exact model ID and rate |
The key insight: the per-token rate is only half the story. Thinking mode generates far more tokens per task as it reasons, so even at a similar rate it costs more per result. For routine work, a faster, leaner variant is cheaper twice over — lower token count and quicker.
Cache-hit discounts
Moonshot discounts cached input tokens. When an agent session resends the same context — system prompt, loaded files — those repeated tokens bill at the cache rate, not full price. Since Claude Code, OpenCode, and Aider all resend context heavily, cache hits meaningfully cut real bills below the headline rate.
Why providers show different prices
Kimi is served by Moonshot directly and by various third-party providers and aggregators, each with its own blended price, speed, and rate limits. If you reach Kimi through an aggregator rather than Moonshot, expect a different rate and throughput. For the canonical price, use Moonshot’s own pricing page; for convenience or failover, an aggregator may be worth a small premium — see OpenRouter vs direct API.
How to estimate your cost
- Estimate tokens per task — remember Thinking inflates output significantly.
- Multiply by the per-million rates for your chosen variant.
- Discount repeated input for cache hits.
- Multiply by tasks per day.
For most individual coders using a fast variant, real Kimi spend stays low — and reserving Thinking for hard tasks keeps it that way.
Before you budget Kimi
- Check current per-million rates on Moonshot
- Note that Thinking inflates token count per task
- Account for cache-hit discounts on repeated context
- Compare Moonshot direct vs any aggregator you use
- Default to a fast variant; escalate to Thinking sparingly
Wrapping up
Kimi K2 is pay-per-token with no flat plan, and the real cost is driven as much by which variant you use as by the rate — Thinking burns more tokens per task than Turbo. Cache-hit discounts cut agent-session bills, and prices vary by provider, so use Moonshot’s page as the source of truth. Default to a fast variant and reserve Thinking for hard problems to keep spend low.
To compare against flat-rate options, see coding plans vs pay-per-token and cheapest AI coding API in 2026.
