方法論

Model catalog and list prices (9 fields: input, output, cache read/write, reasoning, image, audio, web search, per request), context windows, modalities, supported parameters, deprecation dates. Per-provider offers from the endpoints API.

Benchmark scores and model metadata such as training compute, release dates and organizations.

Artificial Analysis, LMArena, LiveBench, Aider, SWE-bench, Scale SEAL, HELM, OpenCompass, and CJK and multilingual suites. Scores are fetched from each leaderboard and range-checked before they are written.

Retired by Hugging Face. BenchGecko keeps its last snapshot; it is no longer fetched.

MCP server listings with package info, transport types, and repository links.

Mindshare signals: daily Hacker News mentions (Algolia API), Wikipedia page views (Wikimedia API) and GitHub stars for every tracked model, lab, agent, person and topic.

Public company figures from filings and the daily market close. Private company figures are researched twice a week and written only when a trusted source page contains the number.

GPU rental prices: median, 10th percentile and minimum on-demand price per GPU-hour for tracked AI chips.

Model metadata for open models: parameter counts, licenses, architectures, downloads.

Chips, memory, foundries, systems and energy, curated from datasheets, earnings calls, filings and utility disclosures. Every record carries an as-of date and a source link.

BenchGecko's own automated tests on the newest models: identity, world map, refusals, knowledge cutoff, tokenizer cost, provider quality, drift. Raw answers are published.