Real-world test: Kimi K3.1 Pro costs 5.4x less than Opus 5.5 per task despite 2x list-price gap
Arindam_1729 · x · 2026-11-12
Measuring actual task costs instead of list prices, Kimi K3.1 Pro runs 5.4x cheaper than Opus 5.5 despite the official pricing suggesting only a 2x gap ($4 vs $2 per 1M input). Same 12 classification and 12 NER tasks, same prompts: Kimi on Nebius Token Factory, Opus via Claude Code; both scored 12/12. Per 1,000 tasks: classification $0.93 vs $5.36 (5.8x), NER $2.12 vs $7.19 (3.4x). Key cost drivers: (1) you pay for unseen tokens — reasoning was Kimi's biggest cost line (35-45% of the bill) even for one-word answers; (2) your harness is part of the bill — 93% of Opus's classification cost came from Claude Code writing 620 tokens to a 1-hour prompt cache.
More from Infra
- AMD runs MolmoAct2 robot policy locally on a Ryzen AI Max+ 395 mini PC — DJiafei · 2026-11-12
- Opinion: AI infra is becoming a coding runtime as intelligence units outgrow tokens — prateekj · 2026-11-12
- PrefixRelay: open-source proxy gives Cursor 97%+ KV cache hits on local models — jjusko20 · 2026-11-12
- Kinetic, a Rust inference engine, answers questions on a 22.6k-token doc in 0.13s vs llama.cpp's 10.95s — jasonkneen · 2026-11-12
- HP starts selling $132,000 GB300 desktop with 748GB memory for local AI — gnukeith · 2026-11-12
- Dev runs out of GitHub Actions minutes, shows off Mac Studio with unreleased M5 Ultra chip — ow · 2026-11-12