Muse Ember, Claude Haiku 5.5, GPT-6 Luna benchmarked

The author benchmarked three small AI models—Meta Muse Ember, Anthropic Claude Haiku 5.5, and OpenAI GPT-6 Luna—on a real-world app tagging task across 200 blog posts using OpenRouter. All three models achieved a nearly identical accuracy between 74% and 79% with zero invalid JSON outputs, but their billing costs and execution speeds varied significantly. The author concludes that OpenAI GPT-6 Luna offers the best balance of low cost and accuracy for this specific task. Open article →