There are countless useful insights in OpenRouter data, which we make public to the world to improve everyone's use of intelligent models. DM me if you're obsessed with data and LLMs and want to be a part of something exciting coming up!

TLDR: Reasoning models → keep detail on auto or high, control cost with reasoning effort Non-reasoning models → low detail genuinely saves money and latency if you can absorb the accuracy hit Full benchmark with per-model tables and methodology:…

So what actually moves the bill? Reasoning effort, by a wide margin. Detail swung accuracy 2 to 17 points while barely touching cost, while capping reasoning effort cut cost 50 to 75 percent with accuracy moving 1 to 2 points, within noise.

OpenAI's low setting downscales everything to 512x512, so those models lost 10 to 17 points. Gemini's low keeps roughly 273 tokens per image part, so both Geminis lost under 3, and gemini-3.1-pro's low run actually came out cheaper.

Low detail for downsamples images, and when the model can't read the fine print it compensates by thinking harder. Net result: more expensive and less accurate at the same time.