Do not buy a chatbot because somebody says it uses the “smartest model.” A website assistant usually needs discipline, speed, and good source material more than frontier-level reasoning.

The current shortlist

As of July 2026, the strongest general-purpose model families for this job come from OpenAI, Anthropic, and Google. Their catalogs change quickly, so production systems should pin a stable model version and test upgrades before switching.

FamilyBest fitOur take
OpenAI GPT-5.6 Terra / LunaTool use, structured actions, balanced cost

Terra is the practical default; Luna suits simpler, high-volume conversations.

Claude Sonnet 5Nuanced language and instruction following

Strong for detailed service explanations and complex policies; test latency and token cost.

Gemini 3.5 FlashFast multimodal and high-volume workflows

Attractive when Google tooling, large context, or mixed media already matters.

Official model references: OpenAI Anthropic Google

Why “best model” is only 20% of the answer

  1. Retrieval: The assistant needs the right business information at the right time. Dumping an entire website into every prompt is slow and unreliable.
  2. Guardrails: It needs explicit boundaries for pricing, promises, medical or legal topics, private data, and anything that requires a person.
  3. Tools: A useful chatbot may need to create a lead, check a service area, request an appointment, or notify staff—not merely generate prose.
  4. Latency: A brilliant answer that takes 18 seconds feels broken. Website visitors expect the first useful response quickly.
  5. Evaluation: You need a repeatable set of real customer questions to test accuracy before and after every model change.

Which tier should you use?

Fast / economical tier

Best for FAQs, routing, lead capture, and tightly defined workflows. This should be the starting point for most small-business websites.

Balanced tier

Use when questions require comparing several services, interpreting longer policies, or reliably calling multiple tools.

Frontier tier

Reserve it for genuinely complex analysis. It is often slower and more expensive than a public website conversation justifies.

Our default recommendation

Start with a stable, lower-cost model in the provider ecosystem that best fits your integrations. Build the knowledge base, guardrails, handoff, and evaluation set first. Move up a tier only when tests prove the cheaper model cannot do the job.

Never use the “latest” alias blindly in production.A provider can change the model behind an alias. Pin a version, evaluate the replacement, then upgrade deliberately.

The buying checklist

  • Which exact model and pinned version will run the chatbot?
  • Where is business knowledge stored and how is it updated?
  • What happens when the system is uncertain?
  • Can it take approved actions, or does it only talk?
  • Are conversations retained, and who can access them?
  • How are model changes tested before production?
  • What ongoing review is included?

Hangar 28 evaluates and manages the model layer; End of Infinity connects that system to the website and customer journey.