LLM Ranking Tool

How the LLM Ranking Tool Works

Every model gets a 0 to 100 score on 11 dimensions grouped into reasoning and knowledge, coding and agents, writing and instructions, context and vision, and speed and cost. Your weights set how much each dimension matters, and the composite is the weighted average: score = sum(weight x dimension) / sum(weights). The stacked bar on each row shows how many points each dimension group contributes, so you can see why a model ranks where it does.

Where the dimension scores come from

Quality dimensions are normalized from public benchmark results: SWE-bench Pro and Verified and Terminal-Bench 2.x for coding and agents, GPQA Diamond and the Artificial Analysis Intelligence Index v4.1 for reasoning, FrontierMath Tier 4 for math, MMLU-Pro and SimpleQA Verified for knowledge, IFBench for instruction following, MRCR for long context, MMMU for vision, and LMArena Elo for writing. Each benchmark maps a floor to 0 and a near-frontier ceiling to 100, then the dimension averages its available benchmarks. Labs stopped publishing many classic benchmarks in 2026, so where nothing verifiable exists the dataset pins an editor estimate flagged with a tilde (~). You can inspect every raw number by expanding a row, and you can override anything with the Edit data button.

Cost and speed

Cost uses a blended price per million tokens under your token mix: (in x input price + out x output price) / (in + out). Use cases differ in shape. Retrieval-augmented generation reads far more than it writes (12:1), while drafting writes more than it reads (1:2), and that changes which model is cheapest for you. Prices map $60 per million to 0 and $0.10 per million to 100 on a log scale. Speed maps 10 tokens per second to 0 and 400 to 100, also log scale. Prices are standard pay-as-you-go rates without batch or cache discounts.

Honest caveats

• Benchmarks are gameable and partially contaminated. Treat small score gaps as noise.
• Terminal-Bench values mix versions 2.0 and 2.1 depending on the source, and LMArena Elo comes from mirror snapshots that disagree by 10 to 15 points.
• Speed is output tokens per second and varies by host. Time to first token on reasoning models includes thinking time and can run minutes on hard tasks.
• Reasoning models bill their thinking as output tokens, so real per-task cost runs higher than per-token prices suggest.