> ## Documentation Index
> Fetch the complete documentation index at: https://docs.a2v2.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Choosing a Model
> How to pick the AI model that powers your agent — balancing answer quality, speed, and credit cost.
Every agent is powered by an AI **model** — the engine that actually writes its
answers. A2V2.ai gives you a menu of models to choose from, and the right pick
depends on what your agent does. A support agent answering thousands of routine
questions has different needs than one reasoning through complex policy.
This page is the *why* behind that choice. For the step-by-step (where to set it)
and the full list of models with their exact credit costs, see [Model &
Temperature](/building/model-settings) — this page won't repeat that table.
## The three things you're trading off
Picking a model means balancing three factors. You rarely get all three at once.
* **Quality** — how well the model understands nuance, follows your
[instructions](/building/instructions), reasons across multiple sources, and
writes a clean answer. Heavier models are stronger on hard, multi-step, or
ambiguous questions.
* **Speed** — how fast the answer comes back. Lighter models tend to respond faster,
which matters when a visitor is watching the reply stream in.
* **Credit cost** — each answer consumes [credits](/concepts#credits), and the cost
per message varies by model. A heavier model can cost several times more per reply
than a light one. At low volume that's negligible; at thousands of conversations a
month it adds up.
The art is matching the model to the job: don't pay for reasoning power a simple FAQ
bot won't use, and don't starve a complex agent of the capability it needs.
## How A2V2.ai labels models
In the model selector on the [Sandbox](/testing/sandbox), each model shows its
**credits per message** and may carry one or both badges:
| Badge | Meaning |
| ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **HIPAA** | The model is eligible for use in HIPAA-regulated workflows. Use these if you're handling health-related information — see [Data privacy & isolation](/concepts/data-privacy-and-isolation). |
| **Premium** | A higher-tier model — typically stronger on quality, and often higher in credit cost per message. |
New agents start on **Claude Haiku 4.5**, a fast, low-cost model at **1 credit per
message** — a sensible default for most support and FAQ use cases.
The available models and their credit costs change over time as providers release
new versions. The selector in the app is always the source of truth for what's
available and what it costs — the full reference table lives on [Model &
Temperature](/building/model-settings).
## When to reach for a heavier model
Start light, and step up only when testing shows you need to. Signs a heavier model
is worth the extra credits:
* **Complex, multi-step questions** — the agent has to combine several sources, do
light reasoning, or weigh conditions ("if the customer is on the annual plan
*and* past their renewal date…").
* **Nuanced tone or judgement** — sensitive topics, negotiation, or carefully worded
responses where a clumsy answer costs you.
* **Long or dense source material** — handbooks, contracts, or technical docs where
the agent must synthesize rather than quote.
* **Instructions aren't being followed** — if a light model keeps ignoring parts of
your [instructions](/building/instructions), a stronger model often holds the line
better.
And signs you can stay light (or step *down*):
* **High-volume, repetitive questions** with answers that live cleanly in your
knowledge base — hours, pricing, "where's my order."
* **Speed is the priority** and the questions aren't subtle.
* **Cost control at scale** — a one-credit model across thousands of chats keeps your
[credit](/concepts#credits) burn predictable.
Don't guess — compare. Use the [Playground](/testing/playground) to run the same
real questions against two models side by side. You'll often find a lighter model
answers just as well for your content, at a fraction of the credit cost.
## A starting point by use case
| Your agent mostly… | A reasonable starting point |
| -------------------------------------------- | ----------------------------------------------------- |
| Answers routine FAQs at high volume | A light, low-cost model (the default is a good start) |
| Handles a mix of simple and tricky questions | A mid-tier model; step up only if testing shows gaps |
| Reasons through complex or sensitive topics | A **Premium** model |
| Works with health-related information | A model with the **HIPAA** badge |
These are starting points, not rules. The reliable way to decide is to test with
*your* knowledge base and *your* real questions.
## A few things to keep in mind
* **Model choice affects file uploads.** If you let visitors attach files in chat,
what's supported (PDFs, images, text) depends on the model — the Sandbox tells you
what the selected model accepts.
* **Quality isn't only the model.** A grounded, accurate answer also depends on good
[sources](/knowledge-base/overview), clear [instructions](/building/instructions),
and a sensible **temperature** setting. A heavier model won't rescue a thin
knowledge base — see [How RAG works](/concepts/how-rag-works).
* **You can change your mind anytime.** Switching models takes effect on the agent's
next answers; it doesn't require retraining your knowledge base.
## Related
Set the model, tune temperature, and see the full model reference.
Run the same questions against two models side by side.
What each answer costs and how credits are metered.
What the HIPAA badge does and doesn't mean.