LLM performance rankings

Benchmark sources

Vellum LLM Leaderboard, Artificial Analysis, official model technical reports, and HuggingFace Open LLM Leaderboard. Scores are the latest publicly reported results as of September 2026, refreshed as new evaluations are published.

Performance per dollar

Shown two ways. The raw Perf/$ column is avg_benchmark_score / blended_cost_per_mtok (blended cost = (3×input + 1×output) ÷ 4 per million tokens). The Best Value ranking is driven instead by the comparability-normalized Norm Value: each model's benchmark strength is first converted to a percentile rank against the whole tracked population (avg_method = percentile_rank_vs_population), then divided by the same blended cost — so a model tested only on easy benchmarks can't post an inflated ratio.

A Norm Value is a percentile-derived value figure, not a % accuracy — only the raw Avg Score and per-benchmark columns are accuracy percentages.

Comparability rule

A model's "average" only covers the benchmarks it has actually been tested on, and the tracked benchmarks differ wildly in difficulty (nobody scores above ~65% on Humanity's Last Exam; AIME scores run into the high 90s), so averaging over different subsets isn't apples-to-apples. Every avg-based ranking and the scatter above therefore require at least 3 filled benchmarks and display the benchmark count (n). Single-benchmark tables (GPQA, SWE-Bench, Terminal-Bench, AIME, HLE and the rest) are unaffected — one shared benchmark is directly comparable.

One row per model

Open-weight models are often sold by several hosts at once — GLM-5.2 by Z.AI, Fireworks and Together. A benchmark score belongs to the weights, so all three score identically, and listing them separately would spend three ranking slots on one answer. Every board here shows a model once, at its cheapest host, with the remaining hosts noted on the row; the price reads “from” whenever more than one host sells it. The counts above follow the same rule — a model sold by three hosts counts once.

The one exception is Fastest endpoints. Tokens per second is a property of the host's serving stack rather than of the weights — Z.AI runs GLM-5.2 at a different speed from Fireworks — so that board ranks endpoints, and a model can legitimately appear on it more than once.