14.1 Customizing Prompt Settings and Models

Every custom prompt you build in AI hub > Prompts (from make.powerapps.com or make.powerautomate.com) runs on a generative AI model you can choose and tune. The AB-410 exam expects you to know which levers exist in the prompt builder, where they live, and — most importantly — which tradeoffs each one introduces. Two places matter: the Model dropdown at the top of the prompt builder, and the settings panel reached by selecting the three dots (…) > Settings.

Selecting the underlying model

Select Model at the top of the prompt builder to pick the model that generates answers. As of mid-2026 the lineup is organized into three categories, each with a different licensing rate and behavior profile:

Exam trap: the older GPT-4o mini and GPT-4o names still appear in legacy AI Builder documentation and remain in use in US government regions; on the commercial cloud the current defaults are the GPT-4.1 family. If a question describes a default, cost-effective model, think "the mini model" rather than memorizing a version number.

Models also carry release stages: Experimental (not for production), Preview (not yet recommended for production), no tag (generally available), and Default (the recommended GA model, also used as fallback if your selected model is turned off). Experimental and preview models can show variability in performance, latency, and availability — never the right answer for a production scenario on the exam. Anthropic Claude models count as experimental even though they show no tag, and experimental external models like Grok 4.1 Fast are explicitly not recommended for production use.

Category guidance

• Mini — fast and cost-effective; summarization, information tasks, routine document processing
• General — superior quality for complex or multimodal work; advanced content creation
• Deep — trained for reasoning; data analysis and multi-step problem solving, but slower because the model reasons before responding and bills at the Premium rate

Temperature

The temperature slider (0 to 1, default 0) controls randomness. At 0 the model produces predictable, consistent, conservative output — what you want for extraction, classification, and compliance-sensitive text. At 1 you get diverse, creative responses — appropriate for brainstorming or marketing copy. Temperature influences output but never guarantees it; responses remain probabilistic even at the same setting. Exam trap: the temperature slider is disabled for GPT-5 reasoning models — if a scenario demands both deep reasoning and a tunable temperature, that combination does not exist.

Output format: text versus JSON

Under prompt settings you can set the Output to Text or JSON. Choosing JSON lets you define a custom format by pasting an example JSON structure, and the model's reply is constrained to that shape (generally available since November 2024). Use text output for prose shown to a human; use JSON output whenever downstream Power Fx logic or a cloud flow must read individual fields — it eliminates fragile string parsing and pairs with the Parse JSON action in Power Automate.

Remaining settings

• Record retrieval — how many records are pulled from attached knowledge sources when grounding the prompt
• Include links in the response — adds link citations for retrieved records
• Enable code interpreter — lets the model generate and execute code, useful for calculation-heavy answers
• Content moderation level — Low, Moderate (default), or High; stricter levels filter more potentially harmful content but answer fewer prompts. Exam trap: this slider is available only for managed models — it is unavailable when you select Anthropic or Azure AI Foundry models.

Tradeoffs and when to switch

Switching models is a cost/latency/reliability decision. A support-summarization prompt processing thousands of emails a day belongs on the mini model at the Basic rate; moving it to a Deep model multiplies credit consumption and slows every call. Conversely, a month-end variance-analysis prompt that must reason across large tables justifies the Premium rate and extra latency of a reasoning model. Keep temperature at 0 for anything deterministic, and pin production prompts to GA models only.

Licensing follows the same pattern: prompts run in Power Apps or Power Automate consume AI Builder credits first, then Copilot Credits; prompts run in Copilot Studio consume Copilot Credits only. Rate tier (Basic/Standard/Premium) scales consumption with the model you chose — so model selection is also the primary cost-control lever.