Models and LLM keys
Choosing a model
How to pick a provider and model for an agent: speed, cost, tool calling, long context and structured output.
Concept1 min readUpdated 23 October 2026
SentientOne charges the same for every reply whatever the model, so the choice comes down to quality, speed and what your provider charges for tokens. You can change the model on an agent at any time; callers don't need to change anything.
What matters for an agent
| If you need | Look for |
|---|---|
| Fast first tokens for a chat widget | The smaller, faster tiers: mini, nano, flash or flash-lite models, Claude Haiku, or Groq-hosted models. |
| Reliable tool calling with MCP | Larger general models. An agent may need several rounds (up to 8) to call tools and then answer. |
| Hard reasoning over long documents | The largest models from each provider (the Pro and Opus tiers). |
| Machine-readable output | Set Output format to JSON, which turns on the provider's JSON mode, and describe the shape in the system prompt. |
| Low token cost at volume | Smaller models. Retrieval already narrows the context to the top 5 passages. |
Measure, don't guess
- 1
Collect real questions
Ten to twenty questions your users actually ask, including hard ones.
- 2
Try them in the Playground
Each reply shows its latency, tokens, sources and tool calls.
- 3
Compare in Analytics
Response time and time to first token per agent are on the Analytics page. Create a second agent with the same prompt on another model to compare side by side.
- 4
Set token prices for cost estimates
Enter the provider's input and output price per million tokens under Settings, Advanced, and the Cost tab estimates model spend.
Keep max_tokens sensible. A very high limit doesn't make answers better, and a reply must finish within 120 seconds.
Related
Still stuck?
Email us with what you tried and any request id. A person replies.