FULL STORY

Kimi K3.1 Tops Terminal-Bench, Then Faces Cost Skeptics

Kimi K3.1 Pro topped Terminal-Bench 4 at 65.7% with roughly $4 per task, but the celebration was quickly followed by a community debate over whether serving Kimi K3 costs over 25x more than DeepSeek, a claim that drew pushback.

2026-10-24 ~ 2026-10-25 · 2 episodes · 9 posts

Episode 1 · Kimi K3.1 Tops Terminal-Bench but Cost Charts Draw Criticism (2026-10-24, 4 posts)

Kimi K3.1 Pro topped Terminal-Bench 4 at 65.7% while NVIDIA's Nemotron 3.5 scored near zero, but Moonshot's cost-per-task chart was criticized as visually misleading, and Kimi costs several times more than DeepSeek V4.1 Flash.

Episode 2 · Debate over Kimi K3 serving costs: claims of 25-30x DeepSeek's (2026-10-24, 5 posts)

A community debate erupted over Moonshot Kimi K3's serving economics. teortaxesTex posted multiple bearish assessments: based on a detailed cost breakdown, K3 needs 21-34x DeepSeek V4.1's compute at equal traffic, with serving costs over 30x V4.1 Flash—so unless Moonshot has clever compute arrangements or a strong small Flash model, the K3.1/Kimi subscription would be hard to sustain.

Confirmed

  • teortaxesTex's estimates: 21-34x DeepSeek's compute at equal traffic; a separate Claude-assisted calculation concluded that even with Blackwell compute, Moonshot could achieve DeepSeek-like margins via subscriptions, but serving K3 still requires 25x the compute.
  • A comparison teortaxesTex reshared as "brutal": matching one DeepSeek serving cluster's 155T tokens/day would require 59,000 GB200s ($2.6B) for Moonshot.
  • YouJiacheng rebutted claims of K3's inherent serving inefficiency as "fakenews," citing 21,600 GB200s at equal traffic and K3 vs DeepSeek V4.1 Flash comparisons under EP64 and 77k average context (DeepInfer published values).

Unconfirmed

  • The two sides use different assumptions (EP configuration, context length, traffic scale, GPU counts); no consensus exists on the actual multiple of K3's serving cost versus DeepSeek's.
  • Whether Moonshot has special compute arrangements or a small Flash model remains teortaxesTex's speculation without primary sources.

Why it matters

  • The debate highlights the tension between subscription pricing and heavy serving costs: if K3 inference really costs tens of times more, the Kimi subscription model's sustainability is questionable; if the rebuttal holds, the rumors may be exaggerated.
  • The cluster-scale figures (59,000 GB200s / $2.6B, or YouJiacheng's 21,600-card figure) illustrate the magnitude of infrastructure gaps among top model providers, offering a window into China's LLM competitive landscape.