Models and reasoning
Each Assistant answers with one model, chosen in Model, and thinks before it answers as much as Reasoning allows. Together they set how good, how fast and how costly its answers are. Both are under Assistant Details on the Configure tab, and Preview answers with them before you save, so you can compare settings on your own questions.
The models
Model | Reasoning |
|---|---|
GPT-5.5 | Off to High |
GPT-5.4 | Off to High |
GPT-5.4 Mini | Off to High. A new Assistant starts on this model, with reasoning Off. |
GPT-5.4 Nano | Off to High |
GPT-4o | None: Reasoning has no effect |
GPT-4o Mini | None: Reasoning has no effect |
The last entry, marked (free) | Always off. See The free option. |
Reasoning
Reasoning sets how much the model thinks before it answers: Off, Low, Medium or High. More reasoning can improve hard answers, especially ones that take several steps or several searches of the Knowledge Base. It also makes visitors wait longer, and the thinking is billed as part of the answer.
While a model reasons, Preview and Reverb 2.0 help show a summary of its thinking, and the Embedded Assistant shows that it is waiting for the answer.
Choosing
The Assistant answers | Start with |
|---|---|
On a website, as the Embedded Assistant: products, pricing, support, finding the right page | A light, fast model |
In Reverb 2.0 help, as the Reverb Assistant: procedures, settings, in-depth questions | A larger model with more reasoning |
Complex subjects that lean on Reasoning Fuel, or troubleshooting in several steps | GPT-5.5 at Medium or High |
Simpler subjects that the Knowledge Base covers well | GPT-5.4 Mini |
GPT-4o and GPT-4o Mini suit casual use, not documentation answers.
A larger model isn't always better. With well-written, well-structured knowledge, a smaller model often answers as well, faster and for less. Run your question list on two or three settings and keep the least costly one that passes. See Test answers with a question list.
An Assistant has one model. To run a light model on your website and a larger one in your help, use two Assistants. See Installs.
The free option
The entry marked (free) answers every question at no charge, with reasoning off, a cap on answer length and a smaller search of the Knowledge Base. Reasoning can't be changed while it is chosen. The same model also appears without (free); chosen that way, it is billed like any other model and can reason.
Free answers have a daily limit per organization. Past it, Assistants in deployment pause until 00:00 UTC; Preview isn't affected. Assistants also switch to the free model on their own when the organization's credits run out, and when their daily budget is reached if it is set to keep answering. See The free model.
Cost and speed
An answer costs the model's rates times the length of the question and of the answer, with the model's thinking counted as part of the answer. Larger models and more reasoning cost more per answer and take longer. Each preview answer shows what writing it cost, so compare settings there. See Credits.
When a model is overloaded, the answer can come from a smaller model, and it is billed at that model's rates.