Two years after 'intelligence too cheap to meter', $10/$50 models are the new norm
teortaxesTex · x · 2026-09-21
- Quoting a two-year-old Sam Altman post, "towards intelligence too cheap to meter" ($0.15/$0.60 per million tokens for GPT-4o-mini), teortaxesTex notes that is exactly DeepSeek V4.1 Flash's pricing today — plus better cache costs and retention — so the cheap tier has stayed flat.
- The twist: $10/$50 models are now the industry norm, and output token prices haven't fallen as the manifesto implied.
Takeaway: the "dirt-cheap intelligence" prophecy only held at the bottom tier, while frontier pricing has moved sharply upward.
More from Infra
- $5, 10-minute SFT on Qwen3.6-35B-A3B lifts GPQA +8% and MMLU-Pro +12% — josh_wills · 2026-09-21
- Agent runtime promises billions of agents and 10-20x sandbox density — astralmatrix · 2026-09-21
- SGLang team helps user debug hicache crash, earning community praise — TheZachMueller · 2026-09-21
- M5 Max vs $2500 desktop upgrade: real-world local LLM speed comparison — vitamins1000 · 2026-09-21
- After finding a memcached bug, Yacine says he'll switch to Valkey — yacineMTB · 2026-09-21
- U.S. data center and information hardware spending now exceeds housing investment — Polymarket · 2026-09-21