Choose a model

The model is the reasoning engine behind an agent's responses. It works with the same system prompt and sources, but different models can vary in instruction following, difficult reasoning, latency, and credit cost. Start with the default model, establish a test baseline, and switch only when your own conversations show a meaningful improvement.

Where to choose a model

You can select a model in two places:

• Playground lets you test a model with a candidate system prompt before saving it.
• Settings → General stores the agent's primary model after you select Save Changes.

In Playground, select Save to Agent to persist the current model and system prompt together. Merely choosing a model or selecting Apply & Test changes the test configuration, not the stored agent.

AgentStack currently uses GPT-5.2 as the default model. The Playground model selector groups available models by provider, labels recommended options, and displays the credit cost beside each option. The selector in Settings → General shows model names in a single list. Treat the Playground selector as the current catalog: providers and model versions can change more frequently than this guide.

The saved choice is the primary model. If its provider fails before response content begins, AgentStack may execute a configured fallback model while keeping billing tied to the selected primary model's fixed credit rate.

How to choose

Use these criteria in order:

1. Correctness on your acceptance set. The model must retrieve and apply your policies accurately.
2. Boundary behavior. It should follow the agent's uncertainty and escalation instructions.
3. Consistency. Repeat important prompts and look for unstable answers.
4. Latency. Faster responses matter for high-volume, conversational support.
5. Credits. Choose the least expensive model that reliably meets the first four criteria.

Smaller one-credit models can be a strong fit for direct, well-grounded questions. Higher-credit models are worth testing when requests require more reasoning or careful instruction following. Credit cost does not tell you a model's context capacity, and a higher price is not proof of better performance on your specific workload.

Run a controlled comparison

Open Playground and select Compare Models. Comparison mode sends one shared message to every active panel. Each panel has its own model, temperature, and system prompt, and the interface shows the combined credit cost across active panels.

To compare models fairly:

1. Use the same system prompt in every panel.
2. Set the same temperature.
3. Send the same acceptance prompts.
4. Grade facts, boundary behavior, and completeness with a simple pass/fail rubric.
5. Repeat the hardest prompts.
6. Save the winning panel only if it improves the full set.

Comparison mode supports two to four panels. You can pause a panel to exclude it from shared sends and the displayed cost. Panel configurations are retained locally for the agent in that browser, but the agent's live configuration changes only when you save a panel to the agent.

Understand credits

The credit number shown beside a model is the per-message model charge used by the selector. A synchronized comparison costs the sum of the active panels' model credits, which Playground displays before sending. For example, comparing two active models means one test message invokes both and charges both model costs.

Credits make repeated tests measurable, but do not skip evaluation to save a few test messages. A short acceptance run is cheaper than deploying an agent that gives incorrect customer answers. See Billing and credits for balances, plan allocations, and top-ups.

When a model change will not help

Switching models is unlikely to fix a missing, stale, or contradictory source. It also cannot make an ambiguous system prompt precise. Follow the response-quality workflow first: verify knowledge, remove conflicts, clarify behavior, and then compare models on the remaining hard cases.