Benchmarks on Blackwell Systemshttps://blog.blackwell-systems.com/categories/benchmarks/Recent content in Benchmarks on Blackwell SystemsHugoen-usSun, 03 May 2026 00:00:00 +0000We Measured It: LSP Saves AI Agents 5-34x Tokens vs Grephttps://blog.blackwell-systems.com/posts/agent-lsp-token-savings/Sun, 03 May 2026 00:00:00 +0000https://blog.blackwell-systems.com/posts/agent-lsp-token-savings/We built a reproducible experiment measuring how many tokens AI coding agents consume when navigating code with grep vs LSP. On HashiCorp Consul (319K lines), LSP uses 34x fewer tokens. On a TypeScript rename across 24 files: 1,441x fewer bytes. The experiment covers 4 codebases, 3 languages, 13 tasks covering 7 agent workflows.