Kimi K3 shows why the pelican test is no longer enough

Moonshot AI launched Kimi K3 with 2.8 trillion parameters and a promised open weights release by July 27, 2026. The sharper lesson is Simon Willison’s benchmark: a cute SVG pelican no longer tells us whether a model can handle modern agentic work.

Kimi K3 arrives as a large and expensive Chinese frontier model

Simon Willison summarizes Moonshot AI’s announcement: Kimi K3 is described by the company as its most capable model, with 2.8 trillion parameters, available through the web and API. An open weights release is promised by July 27, 2026.

Moonshot frames it as the first open 3T-class model. Willison cites Artificial Analysis numbers that put Kimi K3 at 1547 Elo on a long-horizon knowledge work evaluation, at 0.94 dollars per task, with 21 percent fewer output tokens than Kimi K2.6. API pricing of 3 dollars per million input tokens and 15 dollars per million output tokens puts it well above older Kimi pricing.

A tiny SVG prompt exposed the cost of max reasoning

Willison did not stop at benchmark tables. Through OpenRouter, he ran his traditional prompt asking for an SVG of a pelican riding a bicycle. The run cost 25 cents because it used 16,658 output tokens, including 13,241 reasoning tokens.

That is the kind of operational signal a vendor table often hides. For engineering and product teams, the score matters less than the cost of an ordinary interaction and whether reasoning effort can be controlled. In this test, Kimi K3 currently exposes only a max effort level.

The pelican no longer measures the capability that matters most

Willison says the pelican benchmark began as a joke and has largely stopped correlating with top model quality. Modern systems are judged by tool use, long conversations and reliable agentic steps.

The pelican still works as hello world. It confirms access, price, valid SVG output and some spatial reasoning. As a capability ranking, it is now too narrow.

Real adoption will be decided inside tool-heavy workflows

For Kimi K3, the next signals are the open weights release, real costs beyond one prompt and support for lower reasoning effort. Independent tests of tool calling and long tasks will matter more.

If the model fails in agentic workflows, a good pelican will not save it. If it works, it will be another sign that Chinese labs are competing at the top edge, not only on price.

Lilith's verdict

The pelican is a charming canary in the mine. It proves the model is breathing, but it will not run the factory floor full of tools.