Procurement decisions often start with a chart, not a spreadsheet. That is the argument at the center of a September 1 essay from Guido Imperiale, a staff software engineer at OpenTeams, who says the widely cited intelligence-versus-cost plot from ArtificialAnalysis distorts the choice facing most buyers of AI models.
ArtificialAnalysis publishes an Intelligence Index built from a curated benchmark suite, then plots each model’s score against what it cost to run those benchmarks. The resulting Pareto frontier is meant to show the cheapest model available at each intelligence level, a useful reference for teams trying to avoid overpaying for capability they do not need.
Imperiale’s complaint centers on the cost axis. ArtificialAnalysis renders it on a logarithmic scale, which is the only way to fit a $0.015 task next to a $3.69 one on a single chart. But a log scale also compresses what a reader perceives: the nearly 250 times gap between the cheapest and priciest models on the plot reads as a modest visual distance rather than the chasm it actually represents, and the tiny price differences among cheap models look almost as large as they do among expensive ones.
He raises two further objections. ArtificialAnalysis prices open-weight models at the developer’s own API rate rather than the lower rate available through resellers like OpenRouter, and it prices models small enough to run on consumer hardware at datacenter rates instead of the electricity cost of running them locally. Imperiale rebuilt the comparison with a linear cost axis, OpenRouter pricing for six open models, and locally measured electricity costs (roughly $0.20 per kilowatt hour, plus a 15 percent margin for overhead) for models under 35 billion parameters.
His rebuilt plot changes the read considerably. He estimates Fable 5.1, scoring 66 on the index, costs about $3.69 per task, while GLM-5.3, scoring 60, costs about $0.49, a 7.5 times difference for a six-point intelligence gap. Further down the curve, GLM-5.3-Flash with reasoning effort turned up scores 55 at around $0.023 a task, about 160 times cheaper than Fable, while Qwen3.8-27B can run locally on an enthusiast’s existing gaming PC for close to $0.015 in electricity. His conclusion: most users, including many engineers, do not need frontier intelligence and would do fine on cheaper Chinese open-weight models.
None of this has been addressed by ArtificialAnalysis, and the site’s methodology may reflect deliberate tradeoffs Imperiale’s essay does not fully engage with, such as consistency of pricing sources across hundreds of models. Worth noting too: OpenTeams builds and sells services around open-source AI infrastructure, so a conclusion favoring cheap open-weight models over frontier proprietary ones lines up with where the company already makes money. That does not make the log-scale critique wrong, but it is the kind of interest a reader should weigh alongside the argument.
The underlying point survives the conflict of interest regardless: a chart’s axis choice is a design decision, and design decisions carry consequences once the chart lands in front of someone approving a model vendor contract. Whoever skims the plot in a budget meeting is often the person who signs the invoice, and a log scale that visually shrinks a 250 times price gap will bias that signature toward the more expensive option almost every time.
Teams building AI-heavy products should pull ArtificialAnalysis’ raw numbers and replot cost on a linear axis before finalizing a model vendor for their next quarter, rather than trusting the frontier’s default framing.
Guido Imperiale detailed the analysis in a September 1, 2026 post on OpenTeams.
