Why no two AI leaderboards agree
BenchLM and BenchLeader name different number ones this week. Both can be right.
On 29 October BenchLM ranked Gemini 4 Carbon first at 88.4, with Claude Opus 5.5 at 86.3. Two days later the BenchLeader Index put Opus 5.5 first at 69.2, with Carbon at 68.9. Neither site has made a mistake. They measure different things in different ways.
What the scores are built from
BenchLM uses a weighted average of eight categories, with agentic work and coding counting most. BenchLeader standardises results already published elsewhere, then averages them by category. Different recipes give different winners.
Why the gap may mean nothing
BenchLeader reports margins of about ±5 on its leading scores. A 0.3 point lead sits well inside that. BenchLM marks Carbon's score as estimated, with a wide possible range. Treat the top few models as a tie.
What to look at instead
Check whether you can use the model at all. Gemini 4 Carbon currently goes only to selected cybersecurity defenders. Then look at price, and test your own task on two or three candidates.