Models
There is no single best model. How to match tiers to tasks, when routing pays off, and why voice is its own problem.
1 Oct 2026 · 6 min read
There is no single best model. In a Dataiku and Harris Poll survey of 600 CIOs, 81% expected to rely on two or more LLM providers in 2026, and 55% had already switched at least once, mainly to cut costs. Model names change every few months, so this article is about how to choose, not a leaderboard.
A router sends each task, or subtask, to the smallest model that handles it. Published results are encouraging but specific to a benchmark or a vendor: the RouteLLM study reported keeping about 95% of GPT-4 quality while cutting cost by more than 85% on one benchmark, and vendors report 80 to 95% savings on conversational workloads. Treat those as upper bounds and measure on your own traffic. Routing pays off when most requests are simple and volume is high. At low volume, separating classification from generation captures much of the gain.
Streaming pipelines, with speech recognition, the model and speech synthesis all streaming, can get under a second end to end. Realtime speech-to-speech models can reach under half a second with more natural prosody, at the cost of control over each component. Aim for under about 800 milliseconds on phone calls, and escalate genuinely complex turns to a person instead of making the caller wait.
Tell us about one process. We'll show you how an agent could handle it, and what to measure to know it's working.