Blog
Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index, comparable to Opus 4.8 and GPT-5.5
Artificial Analysis
2026.07.16
·LinkedIn·by Homin.Lee#Agent#AI#LLM#Multimodal#Open Weights
Key Points
- 1Moonshot AI's Kimi K3, a 2.8T parameter model, achieves strong agentic performance with an Elo of 1668 on GDPval-AA v2 and ranks first on AutomationBench-AA.
- 2The model demonstrates superior token efficiency, using 21% fewer output tokens than its predecessor while improving its Artificial Analysis Intelligence Index score to 57.
- 3Although Kimi K3 exhibits a higher hallucination rate of 51%, it offers a cost-effective alternative to industry leaders like Claude Opus 4.8 and plans are underway to release its weights publicly.
The Kimi K3 model, developed by Moonshot AI, represents a significant advancement in large language model performance, achieving a score of 57 on the Artificial Analysis Intelligence Index. With a parameter count of 2.8T and a 1M token context window, K3 demonstrates substantial improvements over its predecessor, K2.6.
Performance and Benchmarking
Kimi K3 exhibits strong agentic capabilities, evidenced by an Elo rating of 1668 on GDPval-AA v2, outperforming GLM-5.2 (1514) and Claude Opus 4.8 (1600), though trailing Claude Fable 5 (1760). Notably, it secured the top position on AutomationBench-AA with a 53% success rate. In the AA-Briefcase evaluation, it achieved an Elo of 1547, showcasing high analytical quality and rubric adherence. However, the model’s reliability metrics indicate a regression, with the hallucination rate rising from 39% in K2.6 to 51% in K3, despite a marked increase in the AA-Omniscience Index from +6 to +18, driven by an accuracy improvement from 33% to 46%.Efficiency and Methodology
A core methodological achievement of K3 is its increased token efficiency. The model demonstrated a 21% reduction in output token usage compared to K2.6, utilizing approximately 132M tokens versus 166M to complete identical evaluation tasks. This gain in efficiency is complemented by a higher Intelligence Index score, reflecting superior reasoning per token. The cost per task for K3 is 1.80) and competitive with GPT-5.6 Sol ($1.04).Technical Specifications and Availability
- Architecture: 2.8T parameters; native multimodal input (text and images), with text-only output.
- Pricing: First-party API pricing is set at \$3.00 per 1M input tokens and \$15.00 per 1M output tokens, with cached input tokens discounted by 90% to \$0.30 per 1M.
- Weights: While currently accessible only via Moonshot AI’s first-party API, the company has announced intentions to release model weights. Upon release, K3 is expected to become the premier open weights model, significantly outpacing existing benchmarks like GLM-5.2 and DeepSeek V4 Pro, albeit with a larger physical parameter footprint.
The improvement in agentic performance can be summarized by the increase in Elo rating ():
This delta represents a substantial leap in the model's capacity to execute complex, multi-step workflows autonomously compared to previous iterations.