Kimi K3: China Just Shipped the Biggest AI Model Ever, at Sonnet Prices

Moonshot AI just dropped a bomb. On July 16 the Chinese lab unveiled Kimi K3 with 2.8 trillion parameters, the largest open-weight model ever announced, and a 1-million-token context window. The kicker? It is priced like a Claude Sonnet 5, not like the frontier model it is chasing.

What makes K3 different

Size is only half the story. Moonshot built a fresh architecture around Kimi Delta Attention (KDA), which it claims decodes up to 6.3x faster at million-token context, and a LatentMoE design that fires just 16 of 896 experts per token. Translation: a giant model that runs at a fraction of the per-token cost. Open weights land by July 27, and the vLLM team has already merged native support.

Then there is the price. $3 per million input tokens, $15 per million output, exactly Sonnet 5 territory, with cached input slashed 90% to $0.30. An Opus 4.8 run costs nearly double per task.

How it stacks up against Anthropic

On Artificial Analysis independent Intelligence Index, K3 scored 57, nudging ahead of Claude Opus 4.8 (56) but trailing Claude Fable 5 (60). Cost per task tells the real story: $0.94 for K3 versus $1.80 for Opus. It matches Opus on raw capability and buries it on price. Fable 5 still rules Anthropic lineup, and that is the wall K3 has not broken through.

Where K3 stunned everyone was Arena.ai Frontend Code Arena, a live human-preference battle rather than a static benchmark. It rocketed from 18th to 1st with 1679 points, past Fable 5 and GPT-5.6 Sol, posting a 76% win rate against Fable 63%.

How it stacks up against OpenAI

Days earlier, OpenAI shipped the GPT-5.6 family in three tiers: Sol (flagship, $5/$30), Terra (balanced, $2.50/$15) and Luna (fast and cheap, $1/$6). Sol still tops K3 on the Intelligence Index (59), and Moonshot itself concedes a noticeable gap in user experience against Sol and Fable. But on cost per task, K3 ($0.94) sits right next to Sol ($1.04), while burning 21% fewer output tokens than the last Kimi.

What the community actually built with it

This is where it gets fun. One developer generated a CS:GO x Portal clone in three shots, spending roughly 600k tokens and $3.24. Another let K3 run solo for hours across nearly 1M context to build a browser-based DOS emulator. In long-horizon agent tests, people kept it running for 48 hours without a crash, watching it write its own GPU compiler from scratch.

It is not all glory. Ethan Mollick nailed it: very good open-weights model, but not Sol Max or Fable. Developers keep flagging that it is verbose and slow (26 to 28 tokens per second), that it thinks too much, and some admit they crawled back to Opus when coding tasks flopped. Simon Willison hands-on notes confirm the cost of all that over-reasoning, and Artificial Analysis caught a real regression too: the hallucination rate jumped to 51%, up from 39% last generation.

Still, the community verdict rhymes on one point. This is not good for open source anymore. It is competitive with the best the West has locked behind an API, and only weeks behind. The next act drops July 27, when the weights go public.