OpenTeams engineer Guido Imperiale has replotted Artificial Analysis' model intelligence scores against different cost assumptions. His analysis uses a linear cost axis instead of a logarithmic one and substitutes selected third-party API prices for several model developers' own prices. It does not report a new run of the intelligence benchmarks.
For open-weight models, the comparison draws on OpenRouter offers while excluding providers the author considers too slow, unreliable or lacking zero-data-retention policies. For smaller local models, the plotted cost instead estimates electricity from output-token counts, observed decoding speed and additional power draw. A 15% allowance covers input processing and waits for tool calls.
That local calculation prices existing hardware at zero. Imperiale argues that a suitable gaming or work computer may already serve other purposes, but acknowledges that larger models can require substantial additional hardware spending. Electricity-only points therefore are not full ownership-cost estimates.
Some performance points are estimates too: the article extrapolates a GLM-5.3-Flash reasoning setting and an Ornith model's score. Those points should not be confused with directly measured Artificial Analysis results.