Us: Which task?
At Artisans, we regularly take one real task and run the leading models through it. Same prompt, same data, answers checked one by one.
This time: describing and tagging 41 images. 9 models.
Two surprises:
→ A free, anonymous stealth model (Ox Alpha on OpenRouter) almost tied the winner, Gemini 3.7 Flash.
→ Claude Sonnet 5, the most expensive one in the test, came third. At 5x the price of the winner.
The leaderboard says one thing. Your data says another. We trust the second. Scoreboard attached.
Proud of our team.