LLM Leaderboard Tracker

A self-updating record of how large language models rank against each other over time. Every week a GitHub Actions workflow pulls the official LMArena leaderboard dataset, appends any new snapshots to a CSV history, redraws the trend charts, and refreshes the tables below.

Two questions it answers at a glance:

• Who is winning the frontier race? The best-rated model from each of the top labs, tracked week by week.
• How close is open-weight to proprietary? The best open-weight model against the best proprietary model, and the size of the gap between them.

Latest standings

Tracked history: Text: 164 snapshots · WebDev: 93 snapshots · Vision: 112 snapshots

Text Arena

Leaderboard published 2026-09-30 · change vs 2026-09-13.

#	Δ	Model	Lab	License	Rating	95% CI	Votes
1	🆕	gemini-4-argon-high	google	Proprietary	1533	1524–1542	4,942
2	🆕	claude-opus-5.5-high	anthropic	Proprietary	1512	1502–1522	3,932
3	▼2	claude-fable-5.1-max	anthropic	Proprietary	1511	1504–1517	11,241
4	▼2	claude-opus-5-max	anthropic	Proprietary	1506	1502–1512	28,351
5	▼1	claude-opus-4-6-high	anthropic	Proprietary	1504	1500–1507	77,193
6	▼3	claude-opus-5-high	anthropic	Proprietary	1503	1500–1507	58,380
7	▼2	claude-opus-4-6	anthropic	Proprietary	1498	1494–1501	81,769
8	▼2	gemini-3.8-flash-high	google	Proprietary	1496	1491–1501	24,828
9	🆕	claude-fable-5-high	anthropic	Proprietary	1492	1488–1497	37,900
10	🆕	mimo-v2.6-pro	xiaomi	Open (MIT)	1490	1481–1500	4,074

Best open-weight model: mimo-v2.6-pro (xiaomi, MIT) at #10 with 1490.

WebDev Arena

Leaderboard published 2026-10-01 · change vs 2026-09-24.

#	Δ	Model	Lab	License	Rating	95% CI	Votes
1	=	claude-opus-5.5-max	anthropic	Proprietary	1815	1799–1831	2,062
2	=	gpt-6-astra-max	openai	Proprietary	1788	1777–1798	6,123
3	🆕	claude-sonnet-5.5-xhigh	anthropic	Proprietary	1786	1768–1804	1,531
4	🆕	gpt-6.1-sol-max	openai	Proprietary	1758	1741–1775	1,620
5	▼2	claude-fable-5.1-max	anthropic	Proprietary	1749	1740–1759	6,318
6	🆕	claude-sonnet-5.5-high	anthropic	Proprietary	1715	1702–1729	2,527
7	▼3	claude-opus-5-max	anthropic	Proprietary	1695	1688–1702	16,954
8	▼3	gpt-6-sol-max	openai	Proprietary	1689	1678–1700	3,411
9	🆕	gemini-4-argon-high	google	Proprietary	1680	1666–1693	2,422
10	▼4	qwen3.8-max	alibaba	Proprietary	1671	1659–1683	3,454

Best open-weight model: kimi-k3-max (moonshot, Kimi K3 license) at #13 with 1658.

Vision Arena

Leaderboard published 2026-09-28 · change vs 2026-09-13.

#	Δ	Model	Lab	License	Rating	95% CI	Votes
1	🆕	claude-fable-5-high	anthropic	Proprietary	1326	1319–1334	13,483
2	▲1	claude-opus-5-high	anthropic	Proprietary	1321	1314–1328	15,304
3	▼1	claude-fable-5.1-max	anthropic	Proprietary	1319	1309–1329	4,050
4	=	claude-opus-4-7	anthropic	Proprietary	1317	1310–1324	22,948
5	▲2	claude-opus-4-6-high	anthropic	Proprietary	1316	1309–1322	22,350
6	=	qwen3.8-max	alibaba	Proprietary	1314	1306–1321	10,086
7	🆕	gemini-3.7-flash-high	google	Proprietary	1314	1303–1325	3,425
8	▼3	claude-opus-4-7-high	anthropic	Proprietary	1314	1307–1320	22,559
9	=	claude-opus-4-6	anthropic	Proprietary	1313	1307–1319	26,904
10	🆕	gemini-3.8-flash-high	google	Proprietary	1312	1300–1324	2,711

Best open-weight model: glm-5.3-flash (zai, MIT) at #17 with 1304.

How it works

LMArena dataset (Hugging Face) ──► normalize + filter ──► data/<arena>.csv ──► charts/*.png + README

1. Fetch. Downloads the full split of lmarena-ai/leaderboard-dataset for each tracked arena (Text, WebDev, Vision).
2. Normalize. Keeps the overall category and the top 150 models of each snapshot, and maps the slightly different column names used by different arenas onto one schema.
3. Merge. Adds only snapshots that are not in the CSV yet. Because the whole history is checked every run, a missed week fills itself in.
4. Render. Draws two charts per arena, each in a light and a dark version (GitHub shows the one matching your theme).
5. Publish. Rewrites the section above. If LMArena published nothing new, no file changes and no commit is made.

Data

File	Contents
data/text.csv, data/webdev.csv, data/vision.csv	date, model, organization, license, rating, rating_lower, rating_upper, votes, rank
charts/<arena>-frontier.png	Best rating per lab over the last 12 months
charts/<arena>-open-vs-proprietary.png	Best open-weight vs best proprietary model

Open-weight means any license other than Proprietary in the dataset (Apache 2.0, MIT, Llama, Modified MIT, and similar). Rating is LMArena's Elo-style score from human preference votes, with a 95% confidence interval. Models whose intervals overlap are statistically tied.

The CSVs are plain files, so you can load them straight into pandas:

import pandas as pd
df = pd.read_csv("https://raw.githubusercontent.com/arielshakaramiro/llm-leaderboard-tracker/main/data/text.csv")

Why not the Hugging Face Open LLM Leaderboard?

It was retired by Hugging Face in 2025 and its results are archived, so there is no new data to track. LMArena is the actively maintained leaderboard with an official, versioned dataset.

Run it yourself

pip install -r requirements.txt
python tracker.py

Requires Python 3.10+. No API keys needed, the dataset is public.

Customize

Edit config.json:

• arenas: which arenas to track. Other available ones include search, document, text_to_image, image_edit, text_to_video, and agent.
• history_start: how far back the CSV history goes.
• chart_days: the time window shown in the charts.
• frontier_orgs: how many labs appear in the frontier chart.
• org_colors: a fixed color slot (1–8) per lab, so a lab keeps the same color across every chart and every week.

License

Code: MIT. Leaderboard data: © LMArena, distributed under CC BY 4.0 via the Arena leaderboard dataset.