llm3d

Compare AI models by intelligence, price and tokens used in one rotatable 3D chart.

• Intelligence: the Artificial Analysis Intelligence Index.
• Price: cost per Index task.
• Tokens used: output tokens per Index task.

Live site: https://eomcaleb.github.io/llm3d/

Features

• Pick models by company: each company has a dropdown listing its model families. The chart starts with the top model from a set of leading labs; every other company starts switched off.
• Effort levels: a model family with several effort levels (low → max) is drawn as one line.
• Views: switch between the 3D overview and two flat views, "Intelligence vs price" and "Intelligence vs tokens". Auto-rotate loops a path: it swings one way, swings back the other way, then rises to a top-down view. It's on by default unless reduced motion is enabled; the rotate icon inside the chart turns it off, and dragging or picking a view stops it.
• Efficient-scaling target (on by default): a reference curve for a model that reaches a perfect 100 using only part of the budget. Its dropdown switches it on or off and sets how it climbs and how much budget it needs. Faint connectors show each model's distance from the curve, from green (close) to red (far).
• How the frontier moved: Pareto frontiers of intelligence vs price and intelligence vs tokens for the models released in each quarter, in shades of one color (newer quarters brighter).
• Companies tab: one row per company (Anthropic, OpenAI, SpaceXAI, DeepSeek, Kimi). Each row shows every model family as its own line, across its effort levels, on intelligence vs price and intelligence vs tokens charts. Each model class gets its own color and versions are shades of it; a toggle for each family shows or hides it.

Conventions

CLAUDE.md records how the visualization is meant to work: defaults, colors, what was removed on purpose and the performance rules. Read it before adding models or changing the charts.

Run locally

Open index.html directly. The data ships as data/models.js, so the page works from disk too. Or serve the folder:

python -m http.server 8000
# open http://localhost:8000

Refresh the data

python scripts/build_data.py            # models listed in data/Intelligence Index by Model Creator.xlsx (needs openpyxl)
python scripts/build_data.py --all      # every model Artificial Analysis lists

The spreadsheet decides which models appear and their Intelligence Index scores; release dates, cost, output tokens and time per task come from each model's Artificial Analysis page. Update the spreadsheet (for example with a scheduled task) and rerun the script to refresh the site. A model is only included when Artificial Analysis has measured all four of these:

• Intelligence Index
• Cost per task
• Output tokens per task
• Release date

Models with only an estimated score are skipped.

Deploy (GitHub Pages)

The site is static and served from the repository root on main. The .nojekyll file tells GitHub Pages to serve the files as-is. In Settings → Pages, set the source to "Deploy from a branch", branch main, folder / (root).

Data and attribution

Data: Artificial Analysis. Attribution to Artificial Analysis is required wherever this data is shown. Use of the data is subject to the Artificial Analysis Terms of Use.

License

Code: MIT, see LICENSE.

Performance checks

Run node --test --test-isolation=none tests/scheduling.test.cjs for redraw coalescing, camera backpressure, and deferred chart scheduling checks. Secondary charts render near the viewport; camera animation pauses when the 3D chart is hidden.