Skip to main content
Chatbase lets you choose the AI model that powers each AI agent, from providers such as OpenAI, Anthropic, and Google, or let Chatbase pick for you with Auto. Each model uses a set number of message credits per reply, from 1 to 5 for the recommended models, and offers different strengths in return. This page compares the recommended models by credit cost and use case, so you can balance capability, speed, and cost on every channel your AI agent runs on.

What you can do:

See what you get at each credit tier, and choose the cheapest model that still handles your hardest conversations.
This page covers the recommended models. For the complete list of every available model and its credit cost, see Build → AI Model.

​
At a glance

​
1 credit per message

Auto: Engineered by Chatbase for peak performance, speed, and efficiency. It adapts to every conversation, giving instant answers for everyday questions and frontier-grade intelligence for demanding ones, with multilingual and image understanding built in. One choice that handles every turn end to end, tools included. GPT-6 Luna: The fastest and most cost-efficient tier of OpenAI’s GPT-6 family, optimized for high-volume, low-latency tasks while retaining the core capabilities of the 6 series. Claude 5.5 Haiku: Anthropic’s fastest model, with adaptive thinking and a 1M token context window. It is built for high-volume, latency-sensitive work such as classification, extraction, and routing, at a fraction of the cost of larger models.

​
2 credits per message

GPT-6.1 Sol: OpenAI’s most capable model for complex, multi-step conversations, with support for reasoning, image understanding, and tool use. GPT-6.1 Terra: A highly capable OpenAI model that is also very fast, so your AI agent can handle demanding conversations without slowing down its responses. Claude 5.5 Sonnet: The latest model in Anthropic’s Sonnet series, with strong reasoning, reliable instruction following, and tool use. It offers an excellent balance of intelligence and speed for everyday workflows. Gemini 3.8 Flash: Google’s most intelligent Flash model, built for long-horizon agentic work, software engineering, and complex enterprise workflows. It pairs a 1M token context window with tunable thinking, strong tool use, and image understanding, making it a strong fit for AI agents that need frontier-level reasoning.

​
5 credits per message

Claude 5.5 Opus: The latest and most capable model in Anthropic’s Opus series. It excels at demanding coding challenges, agentic workflows, and sophisticated problem-solving while managing context efficiently.

​
How to choose

Rather than starting from the most capable model, start from the cheapest tier that handles your hardest conversations. Weigh these dimensions against your own traffic:
Don’t judge these from the descriptions alone. Open Build > Instructions and click Compare to run the same message against several models side by side. See Compare area.
Test your hardest real questions against two or three tiers before committing. If the cheaper model handles them, the extra credits buy you nothing.

​
Which model to try for your use case

For most AI agents, start with Auto. If answers are off, check your instructions and Data Sources before switching models, since they affect answer quality more than the model does. Once those are solid, compare other models on your own conversations, as described in Test a model change. The table below lists starting points based on each model’s strengths, not guarantees.

​
Speed, quality, and credits

Every model choice trades between three things:
  • Quality. Higher tiers are built for harder reasoning and longer multi-step tasks. Instructions and Data Sources affect answer quality more than the model does, so fix those first. See Best Practices.
  • Speed. Fast models, such as Auto, GPT-6.1 Sol, GPT-6.1 Terra, and Claude 5.5 Haiku, suit live chat where customers wait for each reply.
  • Credits. The credit cost applies to every reply, so it multiplies with your volume. A reply that triggers an Action uses the model twice, so it costs twice the model’s usual credits.
For example, 1,000 replies cost: On Voice, each reply uses the model’s credits on top of the voice session’s per-minute cost. Phone calls through Twilio or a SIP trunk are billed the same way.

​
Test a model change

Check a new model on your real conversations, and roll back quickly if it does worse.
1

Compare answers side by side

Go to Build > Instructions and click Compare. Open Settings on each chat box and pick a different model, then send the same message to all of them. Use your hardest real questions, including ones that should trigger an Action. See Compare area.
2

Run your Tests before the change

Run your saved Tests on the current model and note each pass rate.
3

Switch the model and publish

Choose the new model in Model configuration, save, and click Publish. Tests run against your published AI agent, not your draft, so publish before you test.
4

Run the Tests again and watch

Run the same Tests again and compare the pass rates. Test replies use the same number of credits as real replies, so a 5-credit model uses test credits five times faster than a 1-credit model. Review conversations in Activity over the next few days. If results get worse, restore the previous version from Version history.

​
Frequently asked questions

Auto is the default model for new AI agents. It is a model option engineered by Chatbase. It adapts to every conversation, giving instant answers for everyday questions and frontier-grade intelligence for demanding ones, with multilingual and image understanding built in. It uses 1 message credit per message.
Go to Build > Instructions and pick a model from the Model dropdown in the Model configuration panel, then save. The change goes into your draft and reaches customers when you click Publish. See Drafts and versions.
Yes. Pick the old model again and publish, or restore an earlier version from Version history. If the old version used a model your current plan doesn’t include, Chatbase adjusts it and tells you what didn’t apply.
A reply that triggers an Action uses the model twice: once to run the Action and once to write the reply. So it costs twice the model’s usual credits. See Actions credit usage.

​
Next steps

Build

See every available model and its credit cost, and set the one your AI agent uses.

Compare area

Run the same message against several models at once.

Best Practices

Instructions and Data Sources affect answer quality more than the model does.

Usage

Track how many message credits your AI agents are consuming.