Kimi K3 beats Opus 4.8 — and matches GPT-5.6 Sol

kimi released k3, a 2.8 trillion parameter open model utilizing a mixture-of-experts architecture with 896 experts. independent testing by artificial analysis validates k3's performance at 57 on the intelligence index, tying opus 4.8 and rivaling gpt-5.6 sol and claude fable 5, particularly excelling in agentic automationbench tasks. however, the model's $3 input and $15 output pricing marks a significant increase over predecessors, matching western costs like sonnet 5 and ending the era of ultra-cheap chinese inference. while full weights arrive in july, the higher pricing and competitive benchmarks suggest chinese labs are prioritizing quality over raw scale.