"Which model should I use" is the wrong first question. The right one is: what does this specific step in the agent actually need — deep multi-step reasoning, fast structured extraction, or just a classification call? Each of those has a different right answer, and the answer for a $0.002 classification call is never the frontier model.
For multi-step reasoning and tool use, we lean on Claude and GPT-4-class models, since the tool-calling reliability at that tier is what makes an agent trustworthy enough to run unattended. For high-volume, low-complexity steps — intent classification, simple extraction — smaller and cheaper models do the job at a fraction of the cost with no meaningful quality loss.
Vendor lock-in is the risk most people underweight. We build the model-calling layer behind a single interface, so swapping providers — or splitting traffic across two for redundancy — is a config change, not a rewrite. That decision alone has saved clients real money when pricing shifted or a new model became the better fit mid-project.
The honest answer to "which LLM is best" is that it changes every few months, and any studio still married to one vendor a year later either hasn't kept up or built something too brittle to change.