LLM landscape

Cost vs capability

What does frontier capability actually cost to train — and which models give you the most capability per dollar? Each priced model is plotted against its ECI score; the efficiency frontier traces the best value at every price tier.

Data as of
Aug 25, 2026
Models priced

31

On the frontier

10

Capability span

98–162

Cost efficiency

Capability per training dollar

Hover near a point to inspect it, click to pin · upper-left is cheaper and more capable

31 models report a published training cost. Claude Opus 4.6 (assumed cost) stands out — 155 ECI for $57.3M, about 7× cheaper than Grok 4 for near-equal capability. Points on the efficiency frontier deliver the most capability available at their price; everything to their lower-right pays more for the same or less.

90100110120130140150160170$0.1M$1M$10M$100MTraining cost, 2023 USD (log scale)Capability level (ECI score)Llama 2-7BFalcon-40BLLaMA-65BLlama 2-70BClaude 2DeepSeek-V3DeepSeek-V3 (Mar 2025)DeepSeek-R1DeepSeek-R1 (May 2025)Grok 4
Organisation
Marker Open weights Proprietary Assumption (outlined)Efficiency frontier

Training cost is Epoch's estimated amortised hardware cost in 2023 USD; capability is the ECI score. Only models with a published cost appear. The efficiency frontier is the Pareto-optimal envelope over measured-cost models — each frontier model offers more capability than any cheaper one — so models with assumed (outlined) costs plot but do not bend the frontier. The featured value pick is drawn from all priced models and is tagged assumed cost when its price is an estimate rather than a measured figure.