Reddit weighs Artificial Analysis against LM Arena

This is the right argument for AI developers to have, because Artificial Analysis and LM Arena answer different questions and neither should be treated as a universal truth machine.

• –Artificial Analysis is strongest when you want structured comparisons across intelligence, speed, price, and methodology rather than pure leaderboard vibes
• –LM Arena is still useful for blind preference testing and real-world taste checks, but crowd voting can drift with prompt mix, hype cycles, and sample bias
• –Broken-out subscores are usually more useful than a single headline score when you care about coding, agentic tasks, hallucination rate, or throughput
• –The practical move is to triangulate: use public benchmarks to narrow the field, then run your own evals on your real prompts before standardizing on a model