Model map

Pick by the job

The reasoning is written by hand; the numbers beside it refresh with the data. Hover a job to find it on the map, or a point on the map to find its job.

Every current model at every effort level, score up and cost per task across on a log scale. Numbered dots are the picks, crosses the rows I would skip.

    Rows I would skip

    Each one costs more than it returns next to something else in the table.

      The speed run

      Measured here, not quoted. The bar is what we got through each lab's own command-line tool; the white tick is Artificial Analysis's median for the same model on the vendor API.

      What a subscription buys

      OpenAI publishes enough to do the maths

      Its plans meter in credits with a published rate, so each one converts into finished intelligence-index tasks a week.

      Every row

      One row per model per effort level. Click a row to pin it on the chart, click a second to compare the two.

      Rows from
      Paid via
      Lab
      Limits
      intel at least $ per task at most out tokens at most
      Choose columns

      Which model people like the look of

      Design Arena's website board: human votes on what looks good, which the intelligence index says nothing about.