GPT-5.4 is a premier model and must be enabled by a tenant administrator before it can be used. See Enabling premier models for instructions.
Agent styles
GPT-5.4 always uses the ReAct Core (react_core) agent style, regardless of the style that is set on your agent. You do not need to configure this explicitly as it is applied automatically.
Agent behavior with GPT-5.4 is architecturally similar to GPT-OSS-120B: the same instruction patterns, tool usage rules, and prompt engineering practices apply. If you have agents that are already tuned for GPT-OSS-120B, they are a solid starting point.
Model behavior is not identical across versions. When you switch to GPT-5.4, you might find that some instructions must be tightened, softened, or reworded to get the same results.
Tool name constraints
GPT-5.4 is served through Azure and enforces the OpenAI function object specification. One constraint that may require changes when you move from GPT-OSS-120B on Groq is the tool name format. watsonx Orchestrate uses the display name of a tool when it builds the function payload that is sent to the model. The display name must therefore conform to the Azure specification:
Display names must match ^[a-zA-Z0-9_-]{1,64}$ — only letters, numbers, underscores, and hyphens, with a maximum length of 64 characters.
This applies to the display names of all tool types that are registered with the agent:
- Skills and tools: The skill display name that is sent in the LLM payload.
- Knowledge bases: The display name that is used to identify the retrieval tool.
- Collaborator agents: The display name that the model uses to invoke a subagent.
Prompt guide
GPT-5.4 uses the same ReAct Core (react_core) agent style as GPT-OSS-120B, so the same prompt engineering principles apply. For instruction templates, tool usage rules, and best practices, see the GPT-OSS-120B prompt guide.