Kimi K3
Moonshot · released 2026-03-15
Quality (AA Index)80
Input price$0.30
Output price$1.20
Context200K
Throughput150 tok/s
P50 latency0.8s
Side-by-side pricing, capabilities, real-world cost across common scenarios, and our editorial pick.
Moonshot · released 2026-03-15
OpenAI · released 2026-01-15
链接通过我们的 OpenRouter 推荐链接在新标签页打开。 查看推广披露
| Metric | Kimi K3 | GPT-5 mini | Verdict |
|---|---|---|---|
Input price /百万 tokens | $0.30 | $0.25 | GPT-5 mini −17% |
Output price /百万 tokens | $1.20 | $2.00 | Kimi K3 −40% |
Context window max input length | 200K | 400K | GPT-5 mini +2.0× |
AA Quality AA Intelligence Index (0–100) | 80 | 84 | GPT-5 mini +4pt |
Arena Elo LMArena human-pref Elo (800–2000) | — | — | Tied |
Throughput tokens per second | 150 | 280 | GPT-5 mini +87% |
P50 latency first token | 0.8s | 0.3s | GPT-5 mini −63% |
Vision multimodal | — | ||
Function calling tool use | Tied | ||
Reasoning mode chain-of-thought | — |
| Scenario | Kimi K3 | GPT-5 mini |
|---|---|---|
customer support 1000K req · 600/180 tok | $396 | $510 |
chat with docs 300K req · 4000/300 tok | $468 | $480 |
code generation 500K req · 2000/500 tok | $600 | $750 |
voice assistant 600K req · 800/200 tok | $288 | $360 |
GPT-5 family inheritance at a fraction of the cost — strong on routine tasks but visibly weaker on multi-step reasoning. Default choice when GPT-5 is overkill.
Set GPT-5 mini as primary, Kimi K3 as fallback. One key, one bill, automatic failover when GPT-5 mini errors.
$ curl https://api.aipricly.com/v1/chat/completions \
-H "Authorization: Bearer $AIPC_KEY" \
-d '{
"routing": {
"primary": "openai/gpt-5-mini",
"fallback": ["moonshot/kimi-k3"]
},
"messages": [{"role": "user", "content": "..."}]
}'