Every AI Model · Tracked

BenchGecko currently tracks 994 AI models across 267 providers, each scored against up to 128 benchmarks. New models are added continuously and the full dataset refreshes daily.

"Best" depends on the task. For general reasoning we rank by average score across 3+ benchmarks; for coding we surface ELO and SWE-bench specifically; for cost/performance we expose a "cheapest capable" metric. Use the filter bar and column sort to define your own winner, or pick a category from the mini tables below.

Average score is the arithmetic mean of a model's normalized benchmark scores, computed only when the model has at least one public benchmark result. Models with fewer than 3 benchmarks are excluded from the podium to avoid single-score outliers.

Model metadata and pricing come from OpenRouter's public API. Benchmarks are pulled from Epoch AI (CC-BY) and SWE-bench's public leaderboards. ELO ratings come from LMArena. Everything is re-normalized and cross-linked daily. See the methodology page for full provenance.

Yes. All BenchGecko data is licensed CC BY 4.0 — attribution required. Use the "Cite this page" button above for ready-made APA, MLA, AP Style, BibTeX, and HTML embed snippets. The free API tier requires a backlink to benchgecko.ai.

Core model, pricing and benchmark data refreshes every 24 hours. Live status and pricing alerts can fire more frequently when upstream sources change. The "Live" pill on this page lights up when the last refresh was less than an hour ago.

Toggle the "Open Source" filter to see only models with OSS weights available. We currently count permissively — weights-available licenses like Llama, Qwen, DeepSeek, and Mistral Open count as OSS for this filter even when the weights come with usage restrictions.

Check the boxes next to any 2-4 models in the leaderboard above, then click "Compare →" in the pink action bar. You can also navigate directly to /compare/[modelA]-vs-[modelB] for shareable comparisons.