Inconsistent behavior – gpt-5-chat-latest does not resolve {{...}} placeholders in the system prompt on the first request

API Feedback
1

Problem summary

When using gpt-5-chat-latest with system instructions containing placeholders ({{user_name}}, {{repository_name}}, etc.), these placeholders remain unrendered in the first response.
The model behaves as if it performs a deferred (lazy) variable hydration, whereas gpt-4.1 and gpt-5 resolve placeholders immediately during the first generation.

Observed behavior

Minimal request:

{ "model": "gpt-5-chat-latest", 
  "input": [ 
  {"role": "system", "content": "Hi {{user_name}} working on your repository. variable; - {{user_name}} = "Alice"}, 
  {"role": "user", "content": "Hi"} ] 
}

Response (first call):

Hi {{user_name}} great to see you working on {{repository_name}}...

Response (second identical call):

Hi Alice great to see you working on Project for applying GenAI best practices...

Expected behavior

For two identical requests, the output should remain consistent.
Placeholders should either always remain literal (no interpolation at all) or always be resolved immediately.
The API documentation does not mention any deferred placeholder resolution mechanism.

Functional impact

  • Loss of identity and tone in the first conversational turn.

  • Inconsistencies in ephemeral agent architectures (MCP, file_search, multi-model orchestration).

  • Potential non-deterministic behavior in systems relying on v1/responses.

Proposed fix or clarification

  1. Explicitly document the deferred hydration behavior currently observed.

  2. Optionally provide a parameter such as:

"resolve_placeholders": "eager"

or

"lazy_injection": false

Affected version

  • Model: gpt-5-chat-latest

  • Endpoint: v1/responses

  • Date: November 2025

  • Reproduced on: REST API, Playground, MCP server, and file_search tool

2

The only place where variables are officially supported are by “prompts” you have saved on the Responses API, then by using the variables API parameter.

https://platform.openai.com/docs/guides/prompting#create-a-prompt

It seems you have manufactured some nonsense string that is not a “placeholder” to be magically resolved, but rather, just some language that the AI can somewhat understand and piece back together as the intention sometimes.

3

The issue does not stem from the use of parameterized prompts, which I am already familiar with implementing.
It is a more subtle behavior: the problem occurs only when using the gpt-5-chat-latest model. In contrast, it does not appear with models such as gpt-5 or gpt-4.1.

Since all these models; including gpt-5-chat-latest; support remote MCP, I wanted to report to the support team that unexpected behavior may occur when using remote MCP with this specific model: placeholders cause issues only with gpt-5-chat-latest, which could represent a potential vulnerability.

It would therefore be advisable either to document this behavior explicitly or to address the issue to ensure consistent behavior across models.