- benchmarks routing cost-optimization
1,000 questions per model: BIRD caught up to Sonnet
1,000 questions per config, GPT-5 baselines added, BIRD rebuilt. On text-to-SQL, routing now matches Claude Sonnet with a gap of -0.7 points.
Carlo Satta6 min read - benchmarks routing cost-optimization
We ran 200 questions per model. Here is what we found.
Routerly routing policies matched Claude Sonnet 4.6 accuracy on MMLU and HumanEval while cutting costs by up to 69%.
Carlo Satta5 min read - announcement open-source llm-gateway
Introducing Routerly: One Gateway for Every AI Model
Today we open-source Routerly, a self-hosted LLM gateway that routes requests across AI providers using intelligent multi-policy scoring, with no external database required.
Carlo Satta4 min read