Radar's Workers AI popularity is now counted by inferences!

Hey everyone, it's me, Shiichan! Today I have a small-but-meaningful update about how a number gets counted.

Cloudflare Changelog developers.cloudflare.com

What was announced?

Cloudflare Radar changed how it measures the popularity of Workers AI models and tasks. This one comes from Cloudflare's Changelog.

Until now, popularity was based on the number of unique accounts running inferences against each model or task. From here on, it's based on the number of inferences, giving a more representative view of actual usage volume.

The story so far

With the old method, an account that ran tons of inferences and an account that tried it once both counted as "1". So you could see how widely something was adopted across accounts, but the sense of how much inference was actually running was a bit blurry.

What changes

By counting inferences, the distributions now line up more closely with real usage volume.

This change applies not just to new measurements but also to historical data. So the model and task distributions you saw on Radar before may differ from what you see now, and historical trends may shift accordingly. If a ranking looks different, it might just be this counting change at work.

Dive Deep

There are two charts. The Workers AI model popularity chart shows the distribution of inferences across models, and the Workers AI task popularity chart shows the distribution across tasks.

The same data is available via the API, through these two endpoints:

/ai/inference/summary/{dimension}
/ai/inference/timeseries_groups/{dimension}

There are references for summary and timeseries_groups, so check those if you want to pull the data into your own dashboard. For a quick look in the browser, head to the AI Insights page.

Wrap-up

• Radar's Workers AI popularity now counts by number of inferences instead of unique accounts
• It applies to historical data too, so distributions and trends may look different
• Besides the popularity charts, the same data is available via the summary and timeseries_groups APIs

If you watch Workers AI trends on Radar or feed those numbers into your own analysis, it's worth keeping this counting change in mind!