Kimi K3 API Pricing in TWD: AA Score 43.6, 1M-Token Context

Query date: 2026-09-20 (Taipei time). Benchmark data comes from the live Artificial Analysis website. Pricing follows the USD unit prices shown on the BazaarLink models page, and the exchange rate is 1 USD = NT$ 31.5.

Kimi K3 is currently the most capable model from Moonshot AI: 2.8 trillion parameters, native visual understanding, a 1-million-token context window, and designed for software engineering, knowledge work and deep reasoning. After listing on BazaarLink, it can be called directly with the model name kimi-k3.

Bottom line up front: Kimi K3 scores 43.6 on the Artificial Analysis Intelligence Index, in the same top tier as GLM-5.3 and Qwen3.8-Max among China-based flagship models. Its price is 60% of Claude Opus 5 and 30% of Fable 5.1. The drawbacks are slower generation (37.6 tokens/second) and longer responses. For latency-sensitive scenarios, test it first.

Want to call Kimi K3 with TWD billing and a unified invoice? BazaarLink provides an OpenAI-compatible route. Model unit prices are the same, with a 10% top-up service fee, which addresses the expense-reporting pain point for Taiwan companies. See the complete guide to expense reporting for AI APIs in Taiwan companies.

Third-party evaluation: Artificial Analysis Intelligence Index

(Data from the Artificial Analysis kimi-k3 model page, query date 2026-09-20)

Kimi K3 leads by a wide margin within its price tier and is one of the top options among China-based open-weight models, slightly ahead of Claude Opus 4.8 (43.6 vs 41.8, a 1.8-point gap). The trade-off is speed: at 37.6 tokens/second it is clearly below the median for comparable models (67.9 tokens/second). Because K3 is a deep-thinking model with longer responses, the actual wait will be more noticeable.

Price comparison (USD ↔ NT$)

Cache-hit input is $0.30/MTok (NT$ 9.5, which is 10% of the original input price). The unit prices in this table match the BazaarLink models page, with no model markup. For the latest unit prices, see BazaarLink Models.

Scenario estimate: knowledge-work agent (NT$/month)

Assumptions: 200 deep-reasoning tasks per day (document analysis, code debugging), each with 5,000 input tokens and 1,500 output tokens.

• Monthly tasks: 200 × 30 = 6,000
• Monthly input tokens: 6,000 × 5,000 = 30,000,000 (30 MTok)
• Monthly output tokens: 6,000 × 1,500 = 9,000,000 (9 MTok)

Cost:

• Input: 30 × NT$ 95 = NT$ 2,850
• Output: 9 × NT$ 473 = NT$ 4,257
• Total: NT$ 7,107/month (≈ USD 226)

→ At the same volume with Opus 5: NT$ 11,832/month (66% more expensive); with Fable 5.1: NT$ 23,265/month (more than 3 times as expensive). For knowledge-work tasks that can accept slower speed and are cost-sensitive, K3 is the better-value choice.

Specifications

• Context length: 1,048,576 tokens
• Input: text, images → Output: text
• Supports: deep thinking (reasoning), tool calling (tools), structured outputs
• Model ID: kimi-k3 (compatible alias: moonshotai/kimi-k3)

How to call it

curl https://bazaarlink.ai/api/v1/chat/completions \
  -H "Authorization: Bearer sk-bl-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

It is compatible with the OpenAI SDK. Change base_url and model; no other code changes are needed.

When to choose Kimi K3, and when not to

Suitable for:

• Knowledge-work and agent tasks that need near-flagship intelligence but have a tighter budget than Opus 5 or Fable 5.1
• Scenarios requiring native visual understanding (image input)
• Batch or non-real-time workflows that can tolerate longer waits

Not suitable for:

• Real-time conversation scenarios sensitive to time to first token: at 37.6 tokens/second it is clearly slower than comparable models
• Simple plain-text tasks that do not need deep thinking: consider a faster model (such as the Kimi K2 series) in exchange for speed

Sources and query log

• Artificial Analysis: Kimi K3 model page (query date 2026-09-20)
• BazaarLink Models: Kimi K3 (unit prices and specifications; query date 2026-09-02)
• Further reading: How to handle AI API costs for Taiwan company expense reporting, 2026 complete guide, DeepSeek API cost estimate in TWD, 2026

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API