Moonshot cuts off new Kimi K3 signups after 48 hours

Moonshot AI suspended new consumer subscriptions to Kimi K3 on July 20, about 48 hours after launch, because request volume maxed out its compute cluster. Remaining GPUs now go to existing paid subscribers only. TechNode broke the story, and Reuters, SCMP, and Global Times corroborated it, so this isn't a rumor.

The fix splits membership into two products: a general Kimi Membership and a separate Kimi Code Membership for programming. That split is an admission.

Coding tasks make dozens or hundreds of model calls per task, each carrying a large context window, versus one call for a chat turn. Moonshot can't serve that at general-purpose rates.

The same July 20 coverage lines up with two more data points. Anthropic's Bun migration burned 5.9 billion input tokens against 690 million output tokens. OpenAI set Codex's context cap right at its 2x price break.

Three companies, three different responses to the same problem: pricing hasn't caught up to what agentic coding costs to serve.

Two days before the Kimi suspension, Alibaba previewed Qwen3.8-Max at WAIC Shanghai. It's a claimed 2.4 trillion parameters, multimodal, priced at 10% of standard rates through Token Plan and Qoder.

There's no benchmark table, no model card, and no disclosed active-parameter count. For a sparse mixture-of-experts model, that's the one number that says what it actually costs to run, and it's not an oversight.

Simon Willison surfaced an unsealed 2022 email from Sam Altman to OpenAI's board, part of the Musk v. Altman discovery. In it, Altman proposed shipping a GPT-3-class open model to make funding harder for rival labs, not to share the technology.

Open-sourcing as a competitive weapon, on the record. The same week, OpenAI's policy lead called Chinese open models "full AI communism."

Kimi K3's weights reportedly ship July 27. That date matters more than launch day did.