Methodology

This page explains where our data comes from, how we compare AI tools and hardware, how our calculators work, and what we have not tested yet.

Where the data comes from

• Prices and plans: each vendor's own pricing page and help center, read on the date shown on each guide. When a page shows different prices by country, we say which country's prices we show.
• Model specifications: official model cards and the config.json files published on Hugging Face (layers, attention heads, context length), and parameter counts from the published weights.
• Hardware: manufacturers' specification and comparison pages.
• Law and regulation: the texts published in the Official Journal of the European Union and, for Spain, the Boletín Oficial del Estado.
• Software behavior: the project's own documentation, README and command reference.

How we compare

For each comparison we decide the criteria before collecting data, record every figure with its source and date, and present the result as a table you can check. Today our comparisons are based on documentation: prices, limits, context windows, features, licenses and data handling. They do not score answer quality, speed or reliability.

How our calculators work

• VRAM calculator: weights are parameters times bits per weight divided by 8, using the bits per weight that llama.cpp measured for each quantization; the KV cache follows each model's published architecture, including sliding-window and linear-attention layers; we add a 1 GiB margin, which is an assumption. We checked the weights formula against 13 real model files: every estimate fell within 3% of the real size.
• LLM API cost calculator: monthly cost is requests times the token prices each provider publishes, with cached input at the cache-read price and batch discounts where offered. The only tokenizer adjustment applied is the one Anthropic documents for its own models. The calculator stops at 200,000 input tokens per request, below the thresholds where some providers charge long-context rates.

Every calculator runs in your browser, labels example values and is covered by automated tests with known values.

What we have not tested yet

As of October 1, 2026, we have not run hands-on tests of AI assistants, coding tools, image or video generators, or local AI speed on specific hardware. Before we publish any quality or performance ranking, we will describe the test on the page: the date, the versions used, the tasks, how results were scored and the limitations.

What we will not do

• Present vendor marketing claims as verified facts.
• Rank or recommend products based on commissions or advertising.
• Publish ratings of products we have not tested with a documented method.