diff --git a/src/observational_memory/llm.py b/src/observational_memory/llm.py
index 162afb9..309592f 100644
--- a/src/observational_memory/llm.py
+++ b/src/observational_memory/llm.py
@@ -102,9 +102,10 @@ def _call_openai_direct(
     import openai
 
     client = openai.OpenAI()
+    token_limit_arg = _openai_token_limit_arg(model, max_tokens)
     response = client.chat.completions.create(
         model=model,
-        max_tokens=max_tokens,
+        **token_limit_arg,
         messages=[
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_content},
@@ -119,6 +120,13 @@ def _call_openai_direct(
     return str(content)
 
 
+def _openai_token_limit_arg(model: str, max_tokens: int) -> dict[str, int]:
+    normalized = model.lower()
+    if normalized.startswith(("gpt-5", "o1", "o3", "o4")):
+        return {"max_completion_tokens": max_tokens}
+    return {"max_tokens": max_tokens}
+
+
 def _extract_anthropic_text(message: object) -> str:
     content = getattr(message, "content", None)
     if not content:
diff --git a/tests/test_llm.py b/tests/test_llm.py
index 0e055de..70b56ba 100644
--- a/tests/test_llm.py
+++ b/tests/test_llm.py
@@ -1,9 +1,12 @@
 """Tests for LLM provider dispatch."""
 
+import sys
+from types import SimpleNamespace
+
 import pytest
 
 from observational_memory.config import Config
-from observational_memory.llm import compress
+from observational_memory.llm import _call_openai_direct, compress
 
 
 @pytest.fixture(autouse=True)
@@ -112,3 +115,34 @@ def fake_openai(system_prompt, user_content, model, max_tokens, cfg):
         assert False, "Should have raised"
     except RuntimeError as e:
         assert "provider 'openai'" in str(e).lower()
+
+
+@pytest.mark.parametrize(
+    ("model", "expected_token_arg"),
+    [
+        ("gpt-5.4", "max_completion_tokens"),
+        ("gpt-5.2-chat-latest", "max_completion_tokens"),
+        ("o4-mini", "max_completion_tokens"),
+        ("gpt-4o-mini", "max_tokens"),
+    ],
+)
+def test_openai_token_limit_parameter_matches_model_family(monkeypatch, model, expected_token_arg):
+    request = {}
+
+    class FakeCompletions:
+        def create(self, **kwargs):
+            request.update(kwargs)
+            return SimpleNamespace(choices=[SimpleNamespace(message=SimpleNamespace(content="ok"))])
+
+    class FakeOpenAI:
+        def __init__(self):
+            self.chat = SimpleNamespace(completions=FakeCompletions())
+
+    monkeypatch.setitem(sys.modules, "openai", SimpleNamespace(OpenAI=FakeOpenAI))
+
+    result = _call_openai_direct("sys", "user", model, 8, Config())
+
+    assert result == "ok"
+    assert request[expected_token_arg] == 8
+    unexpected_token_arg = "max_tokens" if expected_token_arg == "max_completion_tokens" else "max_completion_tokens"
+    assert unexpected_token_arg not in request