Guides · both · kimi k3 pricing
Kimi K3: official price, OpenRouter's cheaper host, and how to call it
Kimi K3 is Moonshot’s open-weight flagship: 2.8 trillion parameters, a 1 million token context, native vision, and a licence that is not a standard open-source grant. The price story is the part people get wrong, because OpenRouter’s card and Moonshot’s card disagree.
Moonshot’s caching doc: cache-miss input $3, cached input $0.30, output $15, per million tokens. Cache writes are extra: $3 for a 5-minute TTL and $6 for one hour. Amazon Bedrock’s Global CRIS standard tier lists the same $3 / $15. Use those when you mean “the lab’s price”. Use the OpenRouter list only when you mean “a host on OpenRouter”.
On 16 Oct 2026 the cheapest input host on the endpoints feed was not Moonshot, and its output price was far above $15. Cheap input with punishing output is a trap for a coding agent.
Benchmarks, with the harness written down
Moonshot’s model card reports Terminal-Bench 2.1 at 88.3 for Kimi K3 using the Kimi Code harness. The other scores on that table are the best the vendor found across different harnesses, and the comparison models are not the current Opus alias. The chart prints the card’s own names. It is not a fresh run by us.
| Entry | score | Detail |
|---|---|---|
| Kimi K3 (max) | 88.3 | Kimi Code |
| Claude Fable 5 (max, w/ fallback) | 88 | Terminus 2 |
| GPT-5.6 Sol (max) | 88.8 | Codex |
| Claude Opus 4.8 (max) | 84.6 | Terminus 2 |
| GPT-5.5 (xhigh) | 83.4 | Codex |
| GLM-5.2 (max) | 82.7 | Claude Code |
A separate write-up has quoted an Artificial Analysis Terminal-Bench figure of 85.0. That is a different harness and a different date, and we do not have the AA feed (it needs a key). We are not averaging 88.3 and 85.0.
LMArena’s coding category, style control, published 15 Oct 2026, puts Kimi K3 (Max) close to the current Sol row and behind the current Opus row. Rank is not a skill certificate. The vote counts are in the table under the chart on the comparison.
Access
The model card says you can select kimi-k3 on platform.kimi.ai, and that Kimi Code can switch to it with /model. The membership guide describes K3 as included for paid Kimi Code tiers, with a 256k variant that uses less quota. We did not see a “sign-ups paused” line in that guide. If you read that claim elsewhere, check the date.
Weights: moonshotai/Kimi-K3. The licence is summarised in the open-source note. Running 2.8T parameters is not a laptop project. “Open weights” is not “free to serve as an API if you are large”.
| Model | Session | vs Opus | vs Sol | Cache price |
|---|---|---|---|---|
| Kimi K3 | $2.61 | 11% cheaper | 78% dearer | published |
| Anthropic: Claude Opus 5.5 | $2.94 | — | — | published |
| OpenAI: GPT-6.1 Sol | $1.47 | — | — | published |
The percentages in this table compare session cost. The price table's percentages compare blended token price. They answer different questions.
That illustrative session is why we call K3 a quality buy, not a discount buy. Flash-tier Qwen or GLM will be the cheap coding loop. K3 is the model you call when the cheap loop fails and you have already decided the output tokens are worth it.
Related: Kimi versus Claude, hosts outside China.
Questions
Is Kimi K3 cheap?
Not on Moonshot's rate. Output is $15 per million tokens. The OpenRouter list price understates the input cost relative to Moonshot's own endpoint.
Is Kimi Code closed to new users?
We could not confirm a current pause. The membership guide we read still describes K3 inside Kimi Code. Check the guide on the day you subscribe.
Sources
- Kimi context caching prices, accessed 2026-10-16
- Kimi K3 model card, accessed 2026-10-16
- Kimi K3 license, accessed 2026-10-16
- OpenRouter Kimi K3 endpoints, accessed 2026-10-16
- Amazon Bedrock Kimi K3 prices, accessed 2026-10-16