While everyone was watching the Fable 5 drama at Anthropic, a Chinese lab quietly raised the stakes: Moonshot AI released Kimi K2.7-Code — a model that doesn’t try to be the world’s best all-rounder. It wants one thing: to write code. As an agent that works its way through your codebase.
What got shipped
Kimi K2.7-Code is a Mixture-of-Experts model with one trillion parameters — but only 32 billion are active per request (384 experts, a handful firing each time). The context window sits at 256,000 tokens. The weights are on Hugging Face, the license is a modified MIT variant. So: self-host away, use it commercially, only the very largest providers have to route through Moonshot’s paid channel.
The numbers Moonshot ships with are the interesting part: +21.8 percent on its in-house Kimi Code Bench v2 over K2.6, plus roughly 30 percent fewer reasoning tokens for the same task. That’s the metric that matters in practice — fewer tokens means lower cost and faster answers.
The price is the real statement
$0.95 per million input tokens, $4.00 per million output. For comparison: Claude’s Fable 5 sits at $10 and $50. It’s not a fair one-to-one comparison, of course — an Anthropic frontier model does far more than a specialized coding agent. But for the narrow task of “go close out this pull request,” that exact gap gets interesting.
And the timing isn’t coincidence, it’s almost comedy: one day before Anthropic moves agent usage into a separate credit pool — yes, that June 15 change — Moonshot drops an open model aimed squarely at the people now staring at their token bill.
My take
I’m cautious with benchmark promises from labs that build their own benchmarks. “+31.5 percent on MLS Bench Lite” tells me little until I’ve seen it on real code myself. What still makes me pay attention: open weights, long context, aggressive pricing. That’s the combination that creates pressure.
Claude Code stays the more complete package for me — the tooling, the plugins, the whole environment around it. But the fact that an open model even enters this conversation is exactly the kind of development I welcome as a user. Competition keeps prices down and the pace high.
Sources: MarkTechPost: Kimi K2.7-Code, LLM Stats: Kimi K2.7 Code, OpenRouter: moonshotai/kimi-k2.7-code