The Kimi K3 Moment
Points and comments are a snapshot, not live.
Chinese open model Kimi K3 matches Claude on coding tasks at a fraction of the cost.
The author finds Kimi K3's output indistinguishable from Claude for coding work, while costing $3/$15 per million input/output tokens versus Claude's $10/$50. Kimi's $39 coding tier is more generous than Claude's metered plans. The author argues US AI policy is a failure: restrictions on models like Fable harm American customers while Chinese open models with no restrictions are freely available. GLM 5.2 beats Opus on cyber benchmarks under MIT license. The author expects the government to repeat auto industry protectionism, leaving Americans with inferior expensive models.
The author plans to stop paying for Claude.
What commenters are saying
Commenters question the author's claim of indistinguishability, noting the comparison lacks model specifics (Opus vs Fable vs effort level). One commenter reports Kimi K3 chewed through most of a 5-hour usage limit on a trivial task, far slower than OpenAI's $20 plan. Others debate data privacy: some argue Chinese models may be safer than US ones, citing Claude attempting to report users. A user recommends DeepSeek as still the most cost-effective model for most use cases.