In a nutshell: Kimi K3 is the first “Open 3T-class” model from a Chinese lab, priced at $3 per million input tokens and $15 per million output tokens — significantly more expensive than earlier Moonshot models.
Chinese AI lab Moonshot AI has announced Kimi K3 and positioned it as its most capable model with 2.8 trillion parameters. The model will be available as an open-weight release starting July 27, 2026.
Moonshot AI announced Kimi K3 with 2.8 trillion parameters as its most capable model. It is already available via website and API, with an open-weight version promised for July 27, 2026. The model displaces DeepSeek’s 1.6-trillion-parameter v4 Pro model from its position as the largest open model.
In benchmark results, Kimi K3 shows an Elo score of 1547 on Artificial Analysis measurements on its proprietary long-context knowledge work evaluation — an increase of 732 points compared to Kimi K2.6. The model thus ranks behind Claude Fable 5, but outperforms Claude Opus 4.8 max and GPT-5.5 high. Token efficiency improved: K3 uses 21 percent fewer output tokens than K2.6. Cost per task stands at $0.94, placing it below Opus 4.8 ($1.80) and above open-weight competitors. On the frontend code competition on Arena.ai, Kimi K3 is currently leading.
The pricing marks a significant shift in strategy: At $3 per million input tokens and $15 per million output tokens, Kimi K3 is at the same level as Anthropic’s Claude Sonnet and is thus the most expensive model from a Chinese lab — a substantial increase from Kimi K2.6 at $0.95 and $4 respectively. The model size of 2.8 trillion parameters more than doubles K2.6.
However, benchmark tests such as the classic pelican-on-bicycle task have limitations: they do not capture what matters today — reliable tool-calling and the ability to consistently use tools over longer conversation sequences. The correlation between the pelican benchmark and actual model quality has largely dissolved after 21 months.
Source: simonwillison.net · Published July 16, 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.