AI agents, inference speed, and the benchmark problem
Anthropic opened Computer Use, Skills API, and Files API. Mistral released Agentic Search, while xAI made Grok Build available to every plan. These launches point to one practical shift: an agent is expected to search, operate software, create files, and deliver a usable result.
The important unit is no longer a single chat answer. It is a bounded workflow with permissions, checks, and a safe handoff. Claude Platform combines browser, software, and file actions; Mistral uses a multi-step search loop.
Inference speed is now a product feature. LMSYS reports Ling-3.0-flash moving from 288 to 606 tokens/s on four Blackwell GPUs. LFM2.5 DSpark uses speculative decoding to increase throughput.
Public scores are not enough. Dreadnode found a gap between passing cyber tasks and actually solving them. Hugging Face found ASR models reproducing test-set errors. Production teams should measure completion, latency, cost, and safe recovery with hidden and fresh tasks.