LinkedIn publication text
LLM Routers
🚀 LLM routers: what the literature actually shows
I'm sharing a technical paper (full document + 41 slides) on the layer that decides which AI model answers each request, and that also handles authentication, budgets, fallback and auditing.
🟣 Why they exist: output prices differ by 60× to about 100× between models (RouteLLM). Choosing well, request by request, matters for the budget.
🔵 The OpenAI standard (/v1/chat/completions) is the contract: switching models becomes configuration, not a code rewrite.
🟢 Evidence: learned routers cut cost by more than 2× with quality kept (RouteLLM). But in LLMRouterBench (33 models, 21 datasets), many methods converge and several do not beat a simple baseline.
🟠 Governance: NIST AI RMF, ISO/IEC 42001, the EU AI Act and LGPD mapped to gateway controls. Note: the Digital Omnibus postponed the AI Act's high-risk rules to Dec 2027 and Aug 2028.
🔴 Security: real CVEs in gateways (SSRF, RCE, log leakage) and attacks on the routing algorithm itself, which in most cases rerouted nearly 100% of queries on open routers.
My takeaway: the control plane is already mature; the intelligent decision still has to prove its gain against a simple baseline, and latency and adversarial robustness are the gaps.
All sources were audited (43 references).
© Aridio Silva @aridiosilva — Outubro 2026