The LLM is the brain of your agent — it transcribes intent, picks what to say, routes to nodes. Shopbell's managed layer offers four LLMs. Which to pick depends on cost vs. quality vs. latency tradeoffs. This guide covers each and our recommendations by use case.
GPT-4o-mini (default, recommended for most)
Latency: 400-700ms first token. Cost: low. Quality: solid for routine business calls — bookings, FAQs, basic discovery. Best for: most DIY self-serve customers, inbound calls where the conversation stays in the script. Default for every new workflow. Switch only if you have a specific reason.
GPT-4o
Latency: 600-1000ms first token. Cost: 7x GPT-4o-mini. Quality: noticeably better for reasoning, complex routing, nuanced objections. Best for: high-value calls (legal intake, real estate consultation, medical triage), where a wrong answer costs the business $1000+.
Claude 3.5 Sonnet
Latency: 500-900ms first token. Cost: similar to GPT-4o. Quality: better than GPT-4o for long-context conversations, nuanced style matching, compliance-sensitive domains (credit, legal, medical). Best for: Credit Repair Specialist template, legal intake agents, complex multi-turn bookings where the agent needs to track many facts across a long call.
OSS Llama 70B (free-for-org)
Latency: 700-1200ms first token (depends on traffic). Cost: zero — included in your $99/mo, no token charges. Quality: close to GPT-4o-mini for most business conversations. Best for: testing, non-customer-facing dev agents, very-high-volume deployments where token costs on other models would add up.
Recommendations by use case
Dental, salon, HVAC, plumber — GPT-4o-mini.
Credit repair, law firm intake — Claude 3.5 Sonnet.
Real estate consultation, medical triage — GPT-4o.
Heavy outbound campaigns where you call 1000s of leads — OSS Llama 70B.
Frequently Asked Questions
Can I switch LLMs after a workflow is in production?
Yes. In the workflow's Settings, change the LLM dropdown, save. New calls use the new model immediately. Existing recordings stay archived with their original model noted for compliance audits.
What if my LLM is rate-limited mid-call?
The call completes on the current LLM. New calls automatically fall back to the next model in your fallback list (configure in Model Configurations). Default fallback: GPT-4o-mini → OSS Llama 70B. The caller hears no difference.
Do you support hosted Mistral or Gemini?
Yes via BYOK — paste a Mistral or Gemini API key into a Model Configuration. Managed hosting may add native Mistral/Gemini in the future.