Kimi K3 matches Claude on coding for a fraction of the price, one dev argues

A developer reports running the open Kimi K3 model side by side with Claude on everyday coding work and finding no practical difference — comparable output quality and near-identical token usage, contrary to the expectation that an open model would be sloppier or less efficient. The economics, by his account, are lopsided: K3’s API runs $3/$15 per million input/output tokens against Claude’s top-tier $10/$50, and Kimi’s $19 and $39 subscription tiers offer far more agent headroom than similarly priced Claude plans, which he says can exhaust their allowance in a single morning of work.

The author’s sharper complaint is about what he calls fine print. He claims Claude couldn’t sustain Fable access on its $20 plan and quietly downgraded it to Opus, arguing that a plan whose headline model can be switched off for economic reasons was never really selling that model. From there he pivots to a broader thesis that US AI policy has backfired: gating American models, in his telling, only constrains American customers, while a frontier-quality, unrestricted model from a Chinese lab beyond US regulatory reach is a free download away. He cites a Semgrep finding of GLM 5.2 outperforming Claude on cyber benchmarks precisely because the restricted model refuses categories of work the open one simply does.

He extends the argument to GLM 5.2 (MIT-licensed, cheap, and by his claim beating the latest Opus) and contrasts Anthropic’s tight metering with OpenAI’s ability to put its GPT-5.6 flagship on a $20 plan, reading that as evidence of greater runway. The piece closes with a bleak prediction — that Washington will subsidize and protect domestic models the way it did the auto industry, leaving the US as the one market stuck with models that are neither the best nor the cheapest and, in his politically charged framing, tied to the current administration. Until then, he says, he can’t justify continuing to pay for Claude. The post is opinion, and several of its pricing and model claims are asserted rather than independently verified.