The leaderboard is not the latency.

Everyone wants to pick an AI model by looking at one number. OpenRouter makes that easy — their public catalog shows token pricing, context length, and benchmark scores pulled from Artificial Analysis. That's helpful. But it's not a stopwatch.

A coding index isn't time-to-first-token. A million-token context window doesn't mean your request finishes fast. And the price per million tokens tells you nothing about what a developer actually experiences while waiting.

Those are token prices, not CPMs. The coding and agentic scores come from Artificial Analysis. They're useful for picking a model — they don't tell you how long someone stared at a spinner.

The missing metric is human time

For RamenSplit, latency isn't a model-ranking contest. It's the window where someone is already waiting for work they asked for. We care about three things: does the sponsor card only show during a real, viewable delay? Does it disappear when the response arrives? Can we attribute the event without reading the prompt?

How to read the table without fooling yourself

1. Use coding and agentic scores as capability signals. They show quality, not speed.
2. Use token prices as infrastructure inputs. They explain why some model routes are cheaper to subsidize — not what a user earns.
3. Measure wait time in the product. Provider, queue, prompt length, reasoning effort, network — everything affects what a person actually experiences.
4. Publish the observation date. Model catalogs change fast. A scoreboard without a timestamp becomes folklore.

That's the honest line between a leaderboard and a latency business. OpenRouter tells us what the model market offers. RamenSplit cares about what the developer experiences while that market does its thing.