LLM Pricing

Compare the cost and quality of large language models.

🌐 Open the live chart

The cost of LLMs is steadily falling while their quality improves. This project puts both dimensions on one chart so that useful, affordable models are easier to find.

A rough estimate of the cost of an LLM is the cost per million tokens of input, mostly from LLMPriceCheck. (Typically, inputs are the bigger component of the cost, compared to outputs.)

A rough estimate of the quality of an LLM is the ELO score on the LMSYS Leaderboard. (This is like the chess ELO score, but for LLMs, where people compare 2 LLMs on the same task.)

The chart combines:

• Cost: input price per million tokens (CPMI), primarily from LLMPriceCheck.
• Quality: ELO scores from the LMSYS / LMArena leaderboard.

Use the Overall, Coding, and Hard views to compare different quality scores. Search by model name or move the month slider to explore how the market has changed over time.

Some LLMs are "pareto optimal", i.e. there is no LLM better in both cost and quality. These are shown in green 🟢 and are the best LLMs to use.

Some LLMs are "pareto suboptimal", i.e. there is no LLM worse in both cost and quality. These are shown in red 🔴 and are the LLMs to avoid.

Last updated: 26 Aug 2026

Related sources: LiveBench · Artificial Analysis models

Understanding the chart

Models in the green 🟢 Pareto-optimal set are not beaten by another model on both price and quality. They are strong value choices.

Models in the red 🔴 Pareto-suboptimal set are beaten by another model on both dimensions and may be worth avoiding.

The chart is a rough comparison, not a recommendation. Prices, availability, context limits, latency, and task-specific performance can change the result.

Updating the data

Prerequisites

• uv for running the Python scripts
• Playwright, which can either connect to Chrome/Chromium through the DevTools Protocol on localhost:9222 or launch its own browser

Run the update script. By default, download.py first tries to connect to Chrome at localhost:9222; if no CDP browser is available, it automatically launches a headless browser through Playwright:

./update.sh

The script downloads the Overall, Hard, and Coding LMArena leaderboards to temporary files and updates the corresponding columns in data/elo.csv.

To force a specific browser mode for one leaderboard:

# Use an existing Chrome DevTools Protocol session
uv run download.py URL output.tsv --browser cdp

# Launch Chrome/Chromium with Playwright
uv run download.py URL output.tsv --browser launch

Set LLMPRICING_CHROMIUM or pass --executable when Playwright cannot find your Chrome/Chromium installation automatically. On Windows, the downloader also checks C:\Users\<user>\AppData\Local\Chromium\Application\chrome.exe. Install the Playwright browser runtime only if you do not want to use an existing Chrome/Chromium installation:

uv run playwright install chromium

To update one leaderboard manually:

uv run download.py https://lmarena.ai/leaderboard/text file.txt
uv run scripts/update_elo.py file.txt --elo data/elo.csv --column overall

Use the following URL and column for the other leaderboards:

Leaderboard	URL path	Column
Overall	/leaderboard/text	overall
Hard	/leaderboard/text/hard-prompts	hard
Coding	/leaderboard/text/coding	coding

download.py --describe prints the machine-readable command contract.

Updating screenshots

Capture chart screenshots with screenshot.py:

uv run screenshot.py --model gpt,gemini,claude [--force]

Use --dry-run to preview the files that would be created. The default output directory is screenshots/. To create a video from a captured sequence:

ffmpeg -framerate 2 -i screenshots/gpt-%03d.png \
  -c:v libvpx-vp9 -pix_fmt yuva420p screenshots/gpt.webm

Repository layout

Path	Purpose
index.html	Main pricing chart page
intelligence.html	Intelligence comparison page
js/main.js	Chart rendering, filtering, tooltips, and interactions
data/elo.csv	Historical model quality and pricing data
data/narrative.json	Scrollytelling chart content
download.py	Extracts leaderboard data through Chrome DevTools Protocol
scripts/update_elo.py	Merges leaderboard results into the CSV
screenshot.py	Captures chart screenshots with Playwright
update.sh	Updates all three leaderboard views