July 17, 2026
Beak performance, bike drama
Kimi K3, and what we can still learn from the pelican benchmark
AI’s newest giant model dropped, but the comments only cared about the pelican chaos
TLDR: Moonshot AI launched Kimi K3, a powerful new chatbot that claims top-tier performance and comes with a notably higher price tag. But commenters were far more entertained by the pelican-on-a-bicycle test, arguing over hidden prompts, benchmark nonsense, and whether anyone should care about a 25-cent bird drawing at all.
Moonshot AI just unveiled Kimi K3, its biggest and most powerful chatbot yet, and on paper the thing is a monster: huge size, strong test scores, pricey access, and even bragging rights for coding. But the real spectacle wasn’t the launch itself — it was the community immediately spiraling into a glorious side quest about whether a pelican riding a bicycle still means anything at all.
That’s because the article uses a goofy recurring test — asking an AI to draw a pelican on a bike — as a quick way to get a feel for a new model. The result cost 25 cents, which triggered instant comment-section combat. One camp basically said, “Calm down, that’s still cheaper than hiring a human,” while another side got suspicious about hidden costs after noticing bizarre token counts and wondering if Kimi is stuffing in an unseen system prompt. Yes, the bird test somehow turned into a mini conspiracy.
And then came the jokes. One commenter wondered if pelican bike drawings are already in AI training data, while another imagined a future where the prompt literally spits out the genetic code for a real giant bicycle-riding pelican. Meanwhile, benchmark fatigue was everywhere: several readers said they’re increasingly skeptical of flashy scoreboards and care more about whether a model has actual taste, surprise, and usefulness in the real world. So yes, Kimi K3 may be a serious new player — but in the court of public opinion, the true winner was absurd bird discourse.
Key Points
- •Moonshot AI launched Kimi K3, a 2.8 trillion-parameter model available via website and API, with open weights promised by July 27, 2026.
- •The article reports that K3’s benchmark results place it mostly ahead of Claude Opus 4.8 max and GPT-5.5 high, but behind Claude Fable 5 and GPT-5.6 Sol.
- •Artificial Analysis reported a 1547 Elo score for K3 on a private long-horizon knowledge-work evaluation, cost per task of $0.94, and 21% fewer output tokens than K2.6.
- •Kimi K3 is priced at $3 per million input tokens and $15 per million output tokens, making it the most expensive model released by a Chinese AI lab to date according to the article.
- •The article argues that the pelican SVG benchmark is no longer a reliable way to compare modern models because it does not measure agentic tool calling or long-conversation tool reliability.