AI Models in HALO

HALO is built on a multi-model architecture. Rather than relying on a single AI model for everything, you can now configure which model powers each part of your setup — optimising for performance, vendor preference or speed depending on the task.

This article covers model configuration across HALO's core features: agents, tools, and guardrails. Behind the scenes, HALO uses additional models for internal functions like automatic translation and platform-level safety, but those aren't covered here.

Default models

The default agent model has been updated to GPT 5.4 — an upgrade from the previous default of GPT 5.1. GPT 5.4 brings improved reasoning and instruction-following, making it better suited for complex agentic workflows.

For tools and guardrails, Claude Haiku 4.5 remains the default — it is fast, consistent, and well-suited to structured, high-volume tasks.

The Decide with AI step always uses Microsoft Decision-1, a model built for classification and scoring. It has no model setting.

Configuring models per component

You can override the default model at each of the following components:

• Agent — set in Agent Settings, under Model Parameters

• Generate with AI step — set within the step configuration in the Tool Editor

• Prompt agent step — set within the step configuration in the Tool Editor

• Custom Guardrail — set per agent within Agent Settings

This gives you fine-grained control: you might use a more capable model for a complex reasoning agent while keeping Haiku for high-frequency tool steps where speed matters more.

For the current list of available models, check the model selector in HALO Studio — the available options are updated as new models are added to the platform.

Model characteristics

GPT 5.4 is adaptive and conversational — it handles open-ended tasks and ambiguity well, and is purpose-built for agentic work. It follows complex instructions without needing rigid prompt structures and produces responses that feel natural.

Claude Haiku 4.5 is fast and precise — it excels at structured, well-defined tasks where consistency matters, such as extracting or summarizing text, and at any high-volume operation where latency matters.

Microsoft Decision-1 is built for decisions only: answering yes/no questions, picking one option from a list, or scoring on a scale. Because it does not write text, it answers much faster than a general-purpose model, and its answers always match the options you defined. It is hosted in the EU and powers the Decide with AI step. For classification, routing, and scoring, use Decide with AI rather than prompting a general-purpose model.

What this means for you

You can now match the model to the task rather than applying a single model across your entire setup. Use a powerful reasoning model where it matters, and a fast, efficient model where throughput and consistency are the priority.

As the AI landscape evolves, so will HALO. Our architecture is designed to adopt new models as they emerge — ensuring you're always building on a strong foundation.