Kimi K3 (Moonshot) Pricing
All plans and pricing as of 2026-09-05
API (Kimi K3)
- ✓K3 (launched 2026-07-16): 2.8T-param open-weight multimodal reasoning model
- ✓1M token context window
- ✓Reasoning effort currently supports only 'max'
- ✓Live on kimi.com, Kimi app, Moonshot API, and OpenRouter; capacity-limited at launch (frequent 429s)
Self-hosted (Free -- K2.6/K2.7 line)
- ✓K3 WEIGHTS PUBLISHED ~2026-07-26/27 at huggingface.co/moonshotai/Kimi-K3 -- safetensors, 2.8T total params, 104B activated, 1,048,576-token context
- ✓K3 ships under its own **Kimi K3 License**, NOT the Modified MIT used for K2 -- read the license terms before commercial use
- ✓K2.6 + K2.7-Code weights remain on Hugging Face under Modified MIT
- ✓Fine-tuning permitted
API (Moonshot direct, K2.6)
- ✓K2.6: $0.60 in / $2.50 out (Moonshot direct)
- ✓256K context
- ✓Native video input (mp4/mov/avi/webm)
Is Kimi K3 (Moonshot) Worth the Price?
Value Score: 8.5/10
Overall Score: 8.1/10 · Agentic coding workflows, tool-use agents, long-horizon repository work (1M context), and teams who want frontier-tier quality at a fraction of frontier pricing ($15/M output vs Fable 5's $50/M).
Kimi K3 (July 16-17, 2026) vaulted Moonshot from 'best open-weights value' to genuine frontier contention: 2.8T parameters, 1M context, multimodal, ranked best-available on Arena.AI at launch, and priced at $3/$15 per 1M -- a fifth of Anthropic's Fable 5 output rate. The launch rattled markets enough to be called a second DeepSeek shock. The open-weight promise has since been kept: the weights landed on Hugging Face around July 26-27 at 2.8T total / 104B activated, though under a bespoke Kimi K3 License rather than the Modified MIT of the K2 line, so read the terms before commercial use. Early August settled the distribution question too -- K3 is now selectable inside GitHub Copilot (hosted on Fireworks AI) and Perplexity (US-hosted), which lets Western teams use it without touching Moonshot's API. Remaining caveats: vendor benchmark claims still await third-party verification, the Moonshot API is visibly capacity-strained, and at 2.8T parameters self-hosting is datacenter work, not a laptop project. If you want maximum capability per dollar, K3 is the first model to test -- via Copilot or Perplexity if jurisdiction is a concern, via the Moonshot API if it isn't.
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
How Kimi K3 (Moonshot) Pricing Compares
| Tool | Free Tier | Starting Price | Value Score | Overall |
|---|---|---|---|---|
| Kimi K3 (Moonshot)(this tool) | Yes | $3 / $15/per 1M tokens (input/output) | 8.5/10 | 8.1 |
| Qwen (Alibaba) | Yes | $0 | 10/10 | 8.8 |
| MiniMax M3 | Yes | $0 | 9.5/10 | 8.4 |
| Gemma 4 (Google) | Yes | $0 | 10/10 | 8.3 |
| IBM Granite 4.0 | Yes | $0 | 9.5/10 | 8.2 |
| gpt-oss (OpenAI) | Yes | $0 | 10/10 | 8.1 |