The Kimi K3 Moment
stephen.bochinski.dev ● Covered by 7 sources
Kimi K3 matches Claude's coding output at a fraction of the cost. And it looks like US policy is hobbling Claude, not helping it.
Based on reporting by stephen.bochinski.dev — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Running Kimi K3 side by side with Claude on ordinary coding work turns out to be a strange experience, mostly because there is nothing strange to report. Same tasks, same quality, and token counts that land in nearly the same place. That alone undercuts the usual assumption that open models are sloppier or need more tokens to reach a comparable answer. The gap that actually shows up is price: K3's API runs $3 per million input tokens and $15 per million output, against $10 and $50 for Claude's top model. On the subscription side Kimi starts at $19 a month, with a $39 coding tier that outpaces anything Claude offers near that price point, while Claude's own plans are metered tightly enough that a full day of agent work can burn through the allowance before lunch.
Then there is the fine print. Claude couldn't keep Fable access working on its $20 plan, so it got switched off, quietly falling back to Opus instead. A model that can be pulled from a plan whenever the economics get uncomfortable was never really the thing being sold in the first place. Kimi's tiers don't carry that kind of asterisk.
The wider story here is a policy failure. The Fable model got held back by the administration, and what eventually shipped is a hobbled version that won't touch entire categories of work. Meanwhile a frontier-quality model with none of those restrictions is one download away, built by a Chinese lab that sits entirely outside US regulatory reach. Semgrep's cyber benchmarks caught this exact effect: GLM 5.2 outperformed Claude simply because the restricted model refuses the task and the open one doesn't. GLM 5.2, released under an MIT license, beats the latest Opus release on real work without ever claiming to be frontier, and costs a fraction of it. OpenAI went through the same government process with GPT-5.6 and still managed to put its flagship on the $20 plan, which suggests it has room to maneuver that Anthropic simply doesn't.
The likely next chapter is regulation aimed squarely at open source, following the same script used on American carmakers: subsidies, bailouts, and tariffs that produced trucks sold at home and largely ignored everywhere else. Public-private partnerships propping up domestic AI models that only work within US borders, and can't compete abroad, would be a strange outcome for the country that built the leading labs. It would leave American users paying more for something not as good, tied to whichever administration is doling out the support. For now, at least, there isn't a strong argument left for paying for Claude.
My take — AI-written commentary, not fact-checked reporting
This is the same protectionist reflex that turned American cars into a product nobody outside the US wanted to buy, and there's no reason to think it plays out differently for AI. Subsidize and shield a domestic model long enough and it stops needing to be good, because it only has to beat the competition that's been legally kept off the shelf. Meanwhile developers who actually ship code don't care about industrial policy, they care about cost per token and whether the model finishes the job, and right now the open, unrestricted options are winning both. Betting American AI leadership on tariffs and bailouts is a bet that history already answered.
Read more about this at: stephen.bochinski.dev