Variant Bench counts how many active catalogue entries belong to one model family and generation. A single generation can appear in several tiers, sizes, or reasoning configurations. The bar answers a practical naming question: how many different entries are hiding behind this model number?
The trench coat joke survives in the icon. The chart itself groups the variants so that a crowded generation earns one tall bar instead of filling the comparison with nearly identical names.
How variants are grouped and counted
We group entries by their creator, model family, and advertised generation. GPT 6 variants belong together, while GPT 6.1 forms a separate group. Gemini and Claude use their own groups even when a version number happens to match another family's number.
Product tiers and reasoning configurations count as separate variants when the catalogue records them as distinct entries. Retired entries are excluded. Duplicate source IDs count once, and legacy aliases are deduplicated so that a spelling change does not create an extra model.
Finding the strongest scored variant
Hover over a generation's bar to see the variant with the highest available Artificial Analysis Intelligence Index in that group. The comparison uses the scores currently available in the collection. A score of zero remains a valid score; a missing score does not become zero.
Some groups have no available intelligence scores. Those groups still have a variant count, but the chart makes no best-model claim for them. The highest recorded index is a useful reference within a generation, while your own task may depend more on latency, price, or a particular capability.
Comparing a whole model generation
Selecting any variant in the model picker includes its complete active generation in this chart. Selecting two entries from the same generation produces one bar. That lets you compare whole families without manually locating every configuration.
The nightly model collection supplies the entries, so a newly tracked variant can increase a bar and a retired entry can reduce it. A taller bar means that we track more active variants for that generation. It does not establish a larger training run, a broader user base, or better answers.
Explore the other ridiculous benchmarks or compare model performance in our AI benchmarks.