The headline on Moonshot’s Kimi K3 is the size — 2.8 trillion parameters, rounded up to claim the first “open 3T-class model,” with open weights promised by July 27. That’s the number that travels. It’s also the number I care about least when I’m deciding whether a model earns a slot in a production routing table.

The two figures worth reading twice are cost-per-task and token efficiency. Artificial Analysis clocks K3 at roughly $0.94 per task — close to GPT-5.6 Sol and about half of Opus 4.8 — while its token usage on the Intelligence Index dropped meaningfully versus K2.6. Raw intelligence gets you onto the leaderboard; token efficiency is what decides whether a model survives contact with a fleet SLA. A model that’s 3% smarter and 2x chattier loses in any real latency-and-cost budget, and self-reported Elo doesn’t surface that trade at all.

Simon’s other point is the one I’d file under muscle memory more than model progress: the pelican-on-a-bicycle SVG probe still teaches you something the benchmarks don’t. Every team shipping RAG or agents needs its own idiosyncratic held-out probe — something the lab couldn’t have trained toward — because vendor-reported numbers are a marketing surface, not an eval harness. The pelican is silly on purpose, and that’s exactly why it works: nobody optimizes for it.

The HN thread is already relitigating the “is 2.8T really 3T” rounding, which is the least interesting fight available. Here’s the question I’d rather see answered before the weights land: when open 3T-class models cost the same per task as closed frontier models, what exactly are you still paying the frontier premium for?