วิธีที่เราประเมินผล

Podium Score รวม 5 ตารางอันดับเป็นตัวเลขเดียวที่เปรียบเทียบได้

The Podium Score Formula

Podium Score = 0.35·Arena + 0.30·Benchmarks + 0.20·Intelligence + 0.15·LLM Stats

• Arena (35%) — Elo rating from head-to-head human preference battles (Arena.ai, 726 models, vote-weighted with confidence intervals).
• Benchmarks (30%) — the mean of all normalized benchmark results we track for the model (25 benchmarks in total).
• Intelligence Index (20%) — Artificial Analysis' composite intelligence index.
• LLM Stats (15%) — the composite score published by llm-stats.com.

When a source has no value for a model, its weight is redistributed proportionally across the remaining sources — so a model is never penalized for being absent from one leaderboard.

Normalization

Each signal is min-max normalized to a 0–100 scale across all 700 tracked models, so a score of 80 always means "80% of the way between the worst and best observed model". Percentage-based benchmarks are used as-is; Elo, indices and prices are normalized against their observed ranges.

Category Scores

Beyond the overall ranking we maintain 17 category leaderboards, each built from the benchmarks most relevant to the task:

• Coding — SWE-Bench Verified, SWE-Bench Pro, LiveCodeBench, SciCode, Terminal-Bench, HumanEval.
• Math — AIME 2025, FrontierMath, MATH.
• Reasoning — GPQA Diamond, HLE, ARC-AGI-2.
• Agentic — OSWorld, Toolathlon, MCP Atlas, τ²-Bench Retail, Apex Agents.
• Knowledge — SimpleQA, MMLU-Pro, MMMLU.
• Multimodal — MMMU, MMMU-Pro, CharXiv-R, ScreenSpot-Pro.
• Coding Agents — SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench Hard, τ²-Bench Retail.
• Scientific Reasoning — GPQA Diamond, Humanity’s Last Exam, SciCode.
• Multilingual — Multilingual MMLU.
• Writing — IFBench, Arena Elo.
• Chat & Assistants — Arena Elo, IFBench.
• Research — Humanity’s Last Exam, GPQA Diamond, MMLU-Pro.
• Tool Calling — Toolathlon, MCP Atlas, τ²-Bench Retail.
• Long Context — MRCR v2, Context window.
• Speed — Output tokens/s.
• Value — Blended price /1M, Cost per task.
• Open Weights — Podium Score.

Data Sources

All data is aggregated from five public leaderboards:

• Arena.ai — last synced 2026-08-04.
• Artificial Analysis — last synced 2026-08-03.
• LLM Stats — last synced 2026-08-02.
• Vellum Leaderboard — last synced 2026-07-24.
• LLMBase — last synced 2026-08-04.

Update Frequency

We re-sync all sources weekly; major model releases trigger an immediate update cycle. The last full refresh: 2026-09-16.

Limitations

No benchmark is perfect and arena preference does not equal task fitness. Use the per-category leaderboards and the Compare tool for decisions about specific use cases. Preview models are flagged and may move as more votes and results arrive.