Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI

Details

Moonshot AI has released Kimi K3, claiming the title of first open-weight model in the ~3 trillion parameter class. Built on a mixture-of-experts architecture with 896 total experts, the model handles images and video natively and supports a one-million-token context window. Full weights are scheduled for public release by July 27; the API is live now. In Kimi's own benchmarks across 35 tests, K3 finished first roughly seven times and second or third in most others — placing it below Claude Fable 5 and GPT 5.6 Sol but ahead of Opus 4.8, GPT 5.5, and Chinese rival GLM-5.2. Benchmark conditions were not uniform: results used three different agent systems (KimiCode, Claude Code, and Codex) depending on the test. Independent lab Artificial Analysis corroborates the positioning, scoring K3 at 57 on its Intelligence Index — on par with Opus 4.8 but below Fable 5 and GPT-5.6 Sol — while flagging a higher hallucination rate versus prior Kimi models, a meaningful regression for reliability-sensitive workloads. Pricing marks a sharp break from prior Chinese AI norms: $3 per million input tokens, $15 per million output. Per-task costs run roughly $0.94 — comparable to GPT-5.6 Sol and about half the price of Opus 4.8. For teams considering open-weight self-hosting, K3 offers genuine frontier-scale capability at the largest open parameter count available, but the hallucination regression and changed cost structure are both factors worth evaluating before committing to production use.