ethangreeney/ai-analysis 发布前沿模型对比地图

Which AI model should you actually use? One map of every frontier model on the three things that matter: how smart it is, how long you wait, and what a task really costs.

Open the live map →

Read it in a glance

Up is smarter. Right is faster. Color is cost per task, from blue under 10¢ to red over $3, so you can spot smart, fast and cheap at a glance. One click switches the colors to the lab instead. The line is the frontier: models no other model beats on both axes. If a model isn't on the line, something is smarter and faster. Models not timed yet sit in their own lane on the left instead of pretending to be slow.

Flip the axis to Cost to see smart-and-cheap instead, or Timeline to watch the intelligence record get broken release by release.

See why a new release matters

Latest releases gives every launch one line on what it did best: the smartest model yet, 12× cheaper than anything as smart, 1.6× faster than anything as smart, or simply how far it moved past the version it replaces.

Click one and the map draws its own line against the version it replaces (dashed) and the best of everything (light). Each setting on the line is labelled by effort, so you can see which one to pick. Switch to the whole lineup (all of GPT-6, say) with one click.

See how fast it's moving

The timeline answers "how much better is today than a year ago?" in three plain sentences: where last year's best model would rank today, and how much cheaper and faster the same smarts are now. Click any of them to open that exact comparison.

Set a budget

Drag the budget slider left. Everything pricier fades out, and the frontier redraws to show the best you can get for that much.

Compare what you use now

Click any dot, then a second one, or open a release and hit Compare to see it against the version it replaces. The card says it in plain terms: +6.8 intelligence, about the same cost per task. Copy the link and the whole comparison travels with it.

Why these numbers

Most comparison charts use token price and tokens per second. Both mislead for reasoning models, which quietly burn thousands of tokens thinking. This map uses what Artificial Analysis measures end to end: cost per Intelligence Index task and median response time per query. Those are the numbers you feel on the bill and in the wait.

Data refreshes automatically several times an hour, new models appear the day Artificial Analysis publishes them, and every day's snapshot is archived in this repo.

Run it locally

npm install
cp .env.example .env   # add your Artificial Analysis API key
npm run fetch          # pull the latest numbers
npm run dev