Change the AI model used by an agent
Choose an available model for the agent’s workload, then compare answer quality, response time and credit usage.
In this article
Before you begin
Open AI agents and select the agent you want to edit. Check its name and workspace before saving: other agents have their own settings.
Step by step
- 1
Prepare a comparison
Write down representative customer questions and the correct policy/source for each. Include an ambiguous question and an unknown answer.
- 2
Choose a model
Find AI model in general settings and select an offered model. Use the cost information and selected-model usage panel as context.
- 3
Save and start fresh
Save the agent, then start a new test conversation so earlier conversation context does not distort the comparison.
- 4
Compare the outcomes
Check factual accuracy, citations, tone, latency and actual credits for the same prompts. Retain the model that meets the task, not just the most fluent answer.
Compare models on the work your agent actually does
Choose representative tasks rather than relying on a model’s name or one fluent answer. A product assistant may need accurate variant matching; a support assistant may need consistent exception handling and reliable handoff behavior.
Keep the prompt, sources and scenarios stable during comparison. Use a fresh conversation to avoid historical context affecting the result. Compare factual correctness first, then response time, style and actual consumption.
The selected-model usage panel reports that model’s recorded activity in the current billing period. It is not an estimate of the price of your next message, and historical usage does not transfer to a different model when you select it.
- Check exact facts and justified next steps.
- Inspect source/tool evidence and unknown-answer behavior.
- Check the languages and message lengths your visitors use.
- Compare actual credits for similar input/output and conversation context.
- Retest after a model or provider is changed or retired.
Choose a model by the outcome you need, not the brand name.
Each bot has a selected primary model. If that provider reaches a quota or capacity limit, Corthex can retry with a compatible fallback model. The bot editor shows model choices and estimated credit costs. Final credit usage reflects the model used and actual input and output usage, so longer conversations can cost more than the listed short-message base.
Test the exact same prompt set after switching models. Compare correctness, source use, tone and response time—not only whether the reply sounds fluent.
Production choices include Gemini 3.6, 3.7 and 3.8 Flash, GPT-OSS 20B and 120B on Groq, DeepSeek V4 Flash, and OpenAI GPT-5.6 Luna, Terra and Sol plus GPT-6 Luna, GPT-6 Sol, GPT-6.1 Sol and GPT-6 Astra. Preview-only provider models are not offered in the bot editor.
Default model: new bots use Gemini 3.6 Flash unless your workspace selection changes. Older saved model IDs are mapped to supported replacements during the migration window.
| Priority | Choose for | Trade-off |
|---|---|---|
| Fast answers | High-volume support and short factual questions | Lower latency, less room for complex reasoning |
| Balanced | Most RAG support workflows | A pragmatic quality and cost balance |
| Deep reasoning | Nuanced policy, technical or multi-step questions | Higher latency and credit use |
Good to know
Available model choices can change. Follow the live picker rather than hard-coding a model name from an old screenshot.
Switching models does not rewrite past replies. Longer prompts, history and outputs can increase actual usage.
Check the result
Save, start a fresh test conversation and compare the answer with your approved business facts. Settings do not rewrite messages already sent.
Questions & troubleshooting
The selected model is absent
It may have been replaced or be unavailable. Choose from the current list and repeat the comparison.
The same question costs more
Compare input/output size, history length and the actual model used, not only the number of messages.
Rate this guide on this device. Ratings are not sent to Corthex.