Kimi K3.1 Pro built a working 3D parade page in one try, for 68 cents

Moonshot AI released Kimi K3.1 Pro on October 23, a post-trained successor to Kimi K3 with a 1-million-token context window, and it was on OpenRouter and in Vercel's AI Gateway the same day. It costs $2 per million input tokens and $10 per million output, down from K3's $3 and $15, Handy AI reports. A Highspeed version that Moonshot says runs two to three times faster costs double.

Moonshot's headline number is its own: 65.7% on Terminal-Bench 4.0, a hair under Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. The Next Web notes that nobody had checked those figures independently, and Handy AI notes that Moonshot ran most of its coding tests inside Claude Code's own harness. The first outside run is in: Artificial Analysis puts it at 67.2%, first on its board, ahead of Claude Sonnet 5.5 at 63.6% and Claude Opus 5.5 at 59.6%. That is a separate evaluation with its own method, a generic agent harness allowed 500 steps, so it does not mean K3.1 gained on Moonshot's figure.

Moonshot also says it took hacking material out of training before the open release. On its in-house exploit test, a version trained without those cuts solved 90.5% of tasks; the shipped model solves 81%, against 33.3% for K3. Full weights are promised by November 4, Unite.AI reports. Until then it is hosted only.

Benchmarks do not say whether a model can build the kind of page Jon ships, so the evening it came out Jon had his agents run it. They sent four prompts through OpenRouter at default settings, one try each, and the same two page prompts to Kimi K3 for comparison. The hardest was his pick: an event page for the Birchwood Luxury Apartments Puppy Halloween Parade, with a 3D apartment building, a parade of costumed puppies walking a path, a camera that follows the mouse, a pet gallery, award categories and house rules.

What the fleet measured

The whole run cost $1.83. Both K3.1 Pro pages worked on the first try with real 3D scenes. The parade page came back in serif and gold, with lit windows, a lamp-lined courtyard, an "Admit One" card, a run-of-show timeline and "The Paw-litzer Prizes." The 7 PM start time on the page is the model's own invention. One flaw on the simpler page: a floating pumpkin covers part of the hero text.

The pages are live to play with: K3.1 Pro's parade page and K3's.

How to use it

Call moonshotai/kimi-k3.1-pro on OpenRouter or Vercel's AI Gateway; add -highspeed for the faster one at $4 and $20. Reasoning is always on, and reasoning_effort takes low, high or max, with max as the default. Temperature and top-p are fixed, and other values return an error, Orca Router notes.

The catch is time and tokens. At max effort a page took 10 to 13 minutes and 50,000 to 70,000 output tokens. That is under a dollar a page, but it is slow, so it suits a background build, not a chat. Small exact tasks came back in six seconds.

What nobody knows yet

Whether lower effort keeps the page quality and cuts the wait: the fleet ran only the default. And whether the coding scores hold outside Claude Code's harness.

What to watch

November 4, when the weights are due and more providers and outside tests can start.