diff --git a/lib/crewai/src/crewai/llms/_finish_reason_utils.py b/lib/crewai/src/crewai/llms/_finish_reason_utils.py
index e79befcc77..1462b44c84 100644
--- a/lib/crewai/src/crewai/llms/_finish_reason_utils.py
+++ b/lib/crewai/src/crewai/llms/_finish_reason_utils.py
@@ -11,13 +11,86 @@
 
 from __future__ import annotations
 
+import logging
 from typing import Any
 
 
+# Providers spell "I ran out of output budget" differently: OpenAI/Azure Chat use
+# ``length``, Anthropic ``max_tokens``, Bedrock ``max_tokens`` via ``stopReason``,
+# Gemini ``MAX_TOKENS``, and the OpenAI Responses API reports
+# ``incomplete_details.reason == "max_output_tokens"`` rather than a finish reason
+# at all. Compared case-insensitively with separators stripped so one predicate
+# covers all of them.
+_TRUNCATION_REASONS = frozenset(
+    {"length", "maxtokens", "modellength", "maxoutputtokens"}
+)
+
+
 def _as_str(value: Any) -> str | None:
+    """Return ``value`` when it is a string, else ``None``.
+
+    Guards against provider types that are not plain strings (test doubles,
+    protobuf enums) leaking downstream.
+    """
     return value if isinstance(value, str) else None
 
 
+def is_truncated(finish_reason: str | None) -> bool:
+    """Whether ``finish_reason`` means the response was cut off by the token cap.
+
+    Providers spell this differently: ``length`` (OpenAI, Azure), ``max_tokens``
+    (Anthropic, Bedrock) and ``MAX_TOKENS`` (Gemini). Comparison is
+    case-insensitive with separators stripped so one predicate covers all of them.
+
+    Args:
+        finish_reason: The raw finish/stop reason as reported by the provider.
+
+    Returns:
+        ``True`` if the value denotes truncation by the output-token cap.
+    """
+    if not isinstance(finish_reason, str):
+        return False
+    return (
+        finish_reason.replace("_", "").replace("-", "").strip().lower()
+        in _TRUNCATION_REASONS
+    )
+
+
+def warn_if_truncated(
+    finish_reason: str | None,
+    max_tokens: int | float | None = None,
+    model: str | None = None,
+) -> bool:
+    """Log a warning when a response was cut off by the token cap.
+
+    A truncated response is otherwise indistinguishable from a complete one to
+    everything downstream, so the caller cannot tell "the model was interrupted"
+    from "the model answered badly".
+
+    Args:
+        finish_reason: The raw finish/stop reason as reported by the provider.
+        max_tokens: The cap actually sent on the wire, so the message names the
+            setting the reader needs to change. Providers differ here
+            (``max_output_tokens`` on Gemini, ``max_completion_tokens`` on
+            OpenAI and Azure reasoning models), so callers pass their effective
+            value rather than a single shared attribute.
+        model: The model the request was sent to, for diagnostics.
+
+    Returns:
+        ``True`` if a warning was emitted, ``False`` otherwise.
+    """
+    if not is_truncated(finish_reason):
+        return False
+    logging.warning(
+        "Response truncated due to max_tokens limit (finish_reason=%r%s%s). "
+        "The output is incomplete; consider increasing max_tokens.",
+        finish_reason,
+        f", model={model}" if model else "",
+        f", max_tokens={max_tokens}" if max_tokens is not None else "",
+    )
+    return True
+
+
 def extract_choices_finish_reason_and_id(
     response_or_chunk: Any,
 ) -> tuple[str | None, str | None]:
diff --git a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
index 77236249af..a0dfbea52b 100644
--- a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
@@ -9,6 +9,7 @@
 
 from crewai.events.types.llm_events import LLMCallType
 from crewai.hooks.dispatch import HookAborted
+from crewai.llms._finish_reason_utils import warn_if_truncated
 from crewai.llms.base_llm import (
     BaseLLM,
     JsonResponseFormat,
@@ -1055,6 +1056,7 @@ def _handle_completion(
         self._track_token_usage_internal(usage)
 
         finish_reason, response_id = self._extract_finish_reason_and_id(response)
+        warn_if_truncated(finish_reason, self._effective_max_tokens(), self.model)
 
         if _is_pydantic_model_class(response_model) and response.content:
             if use_native_structured_output:
@@ -1603,6 +1605,7 @@ async def _ahandle_completion(
         self._track_token_usage_internal(usage)
 
         finish_reason, response_id = self._extract_finish_reason_and_id(response)
+        warn_if_truncated(finish_reason, self._effective_max_tokens(), self.model)
 
         if _is_pydantic_model_class(response_model) and response.content:
             if use_native_structured_output:
diff --git a/lib/crewai/src/crewai/llms/providers/azure/completion.py b/lib/crewai/src/crewai/llms/providers/azure/completion.py
index 8cdb49cca7..a21611e8ac 100644
--- a/lib/crewai/src/crewai/llms/providers/azure/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/azure/completion.py
@@ -10,7 +10,10 @@
 from typing_extensions import Self
 
 from crewai.hooks.dispatch import HookAborted
-from crewai.llms._finish_reason_utils import extract_choices_finish_reason_and_id
+from crewai.llms._finish_reason_utils import (
+    extract_choices_finish_reason_and_id,
+    warn_if_truncated,
+)
 from crewai.llms.hooks.base import BaseInterceptor
 from crewai.utilities.agent_utils import is_context_length_exceeded
 from crewai.utilities.exceptions.context_window_exceeding_exception import (
@@ -873,6 +876,7 @@ def _process_completion_response(
         self._track_token_usage_internal(usage)
 
         finish_reason, response_id = self._extract_finish_reason_and_id(response)
+        warn_if_truncated(finish_reason, self._effective_max_tokens(), self.model)
 
         # Without available_functions, return tool_calls so the caller (executor) handles execution
         if message.tool_calls and not available_functions:
diff --git a/lib/crewai/src/crewai/llms/providers/gemini/completion.py b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
index 98edc12383..2431593f01 100644
--- a/lib/crewai/src/crewai/llms/providers/gemini/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
@@ -11,6 +11,7 @@
 
 from crewai.events.types.llm_events import LLMCallType
 from crewai.hooks.dispatch import HookAborted
+from crewai.llms._finish_reason_utils import warn_if_truncated
 from crewai.llms.base_llm import BaseLLM, LLMCallBlockedError, llm_call_context
 from crewai.llms.hooks.base import BaseInterceptor
 from crewai.utilities.agent_utils import is_context_length_exceeded
@@ -877,6 +878,7 @@ def _process_response_with_tools(
             Final response content or function call result
         """
         finish_reason, response_id = self._extract_finish_reason_and_id(response)
+        warn_if_truncated(finish_reason, self._effective_max_tokens(), self.model)
 
         if response.candidates and (self.tools or available_functions):
             candidate = response.candidates[0]
diff --git a/lib/crewai/src/crewai/llms/providers/openai/completion.py b/lib/crewai/src/crewai/llms/providers/openai/completion.py
index 20234e5019..4e5ae10f50 100644
--- a/lib/crewai/src/crewai/llms/providers/openai/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/openai/completion.py
@@ -37,7 +37,10 @@
 
 from crewai.events.types.llm_events import LLMCallType
 from crewai.hooks.dispatch import HookAborted
-from crewai.llms._finish_reason_utils import extract_choices_finish_reason_and_id
+from crewai.llms._finish_reason_utils import (
+    extract_choices_finish_reason_and_id,
+    warn_if_truncated,
+)
 from crewai.llms.base_llm import (
     BaseLLM,
     JsonResponseFormat,
@@ -994,6 +997,13 @@ def _handle_responses(
             finish_reason, response_id = self._extract_responses_finish_reason_and_id(
                 response
             )
+            # The Responses API reports ``status`` ("incomplete"), not a finish
+            # reason, so the cause lives in ``incomplete_details.reason``.
+            warn_if_truncated(
+                self._responses_truncation_reason(response) or finish_reason,
+                self._effective_max_tokens(),
+                self.model,
+            )
 
             if self.parse_tool_outputs:
                 parsed_result = self._extract_builtin_tool_outputs(response)
@@ -1141,6 +1151,13 @@ async def _ahandle_responses(
             finish_reason, response_id = self._extract_responses_finish_reason_and_id(
                 response
             )
+            # The Responses API reports ``status`` ("incomplete"), not a finish
+            # reason, so the cause lives in ``incomplete_details.reason``.
+            warn_if_truncated(
+                self._responses_truncation_reason(response) or finish_reason,
+                self._effective_max_tokens(),
+                self.model,
+            )
 
             if self.parse_tool_outputs:
                 parsed_result = self._extract_builtin_tool_outputs(response)
@@ -1921,6 +1938,9 @@ def _handle_completion(
                 )
                 parsed_object = parsed_response.choices[0].message.parsed
                 if parsed_object:
+                    warn_if_truncated(
+                        parsed_finish_reason, self._effective_max_tokens(), self.model
+                    )
                     self._emit_call_completed_event(
                         response=parsed_object.model_dump_json(),
                         call_type=LLMCallType.LLM_CALL,
@@ -1946,6 +1966,7 @@ def _handle_completion(
             finish_reason, response_id = self._extract_chat_finish_reason_and_id(
                 response
             )
+            warn_if_truncated(finish_reason, self._effective_max_tokens(), self.model)
 
             # Without available_functions, return tool_calls so the caller (executor) handles execution
             if message.tool_calls and not available_functions:
@@ -2349,6 +2370,9 @@ async def _ahandle_completion(
                 )
                 parsed_object = parsed_response.choices[0].message.parsed
                 if parsed_object:
+                    warn_if_truncated(
+                        parsed_finish_reason, self._effective_max_tokens(), self.model
+                    )
                     self._emit_call_completed_event(
                         response=parsed_object.model_dump_json(),
                         call_type=LLMCallType.LLM_CALL,
@@ -2374,6 +2398,7 @@ async def _ahandle_completion(
             finish_reason, response_id = self._extract_chat_finish_reason_and_id(
                 response
             )
+            warn_if_truncated(finish_reason, self._effective_max_tokens(), self.model)
 
             # Without available_functions, return tool_calls so the caller (executor) handles execution
             if message.tool_calls and not available_functions:
@@ -2715,6 +2740,22 @@ def _extract_chat_finish_reason_and_id(
         """
         return extract_choices_finish_reason_and_id(response)
 
+    @staticmethod
+    def _responses_truncation_reason(response: Any) -> str | None:
+        """Why a Responses-API call stopped early, when it did.
+
+        ``status`` only says ``"incomplete"``; the cause is carried separately on
+        ``incomplete_details.reason`` (for example ``"max_output_tokens"``).
+        Returns ``None`` for complete responses and for shapes lacking the field.
+        """
+        if getattr(response, "status", None) != "incomplete":
+            return None
+        details = getattr(response, "incomplete_details", None)
+        reason = getattr(details, "reason", None)
+        if reason is None and isinstance(details, dict):
+            reason = details.get("reason")
+        return reason if isinstance(reason, str) else None
+
     @staticmethod
     def _extract_responses_finish_reason_and_id(
         response: Any,
diff --git a/lib/crewai/tests/llms/test_finish_reason_utils.py b/lib/crewai/tests/llms/test_finish_reason_utils.py
new file mode 100644
index 0000000000..bb3dec458d
--- /dev/null
+++ b/lib/crewai/tests/llms/test_finish_reason_utils.py
@@ -0,0 +1,139 @@
+"""Truncation detection shared across providers.
+
+A response cut off by the token cap is otherwise indistinguishable from a complete
+one, so callers cannot tell "the model was interrupted" from "the model answered
+badly". These tests pin the behaviour rather than the wording.
+"""
+
+import logging
+
+import pytest
+
+from crewai.llms._finish_reason_utils import is_truncated, warn_if_truncated
+
+
+@pytest.mark.parametrize(
+    "finish_reason",
+    [
+        "length",  # OpenAI, Azure
+        "max_tokens",  # Anthropic, Bedrock
+        "MAX_TOKENS",  # Gemini
+        "maxTokens",
+        "max-tokens",
+        "  Length  ",
+    ],
+)
+def test_recognises_every_provider_spelling(finish_reason):
+    """Every vendor spells truncation differently and all of them must be caught.
+
+    A provider whose spelling is missed looks identical to a complete response,
+    which is the failure this helper exists to prevent.
+    """
+    assert is_truncated(finish_reason) is True
+
+
+@pytest.mark.parametrize(
+    "finish_reason",
+    ["stop", "STOP", "tool_calls", "content_filter", "end_turn", "", None, 42, object()],
+)
+def test_ignores_non_truncation_reasons(finish_reason):
+    """Normal completions and non-string values must not raise a false alarm.
+
+    A warning on every healthy turn is worse than no warning at all, because
+    readers learn to ignore it.
+    """
+    assert is_truncated(finish_reason) is False
+
+
+def test_warns_once_and_names_the_budget(caplog):
+    """One warning per truncated response, naming the cap that caused it.
+
+    The cap has to appear or the reader cannot tell which setting to raise.
+    """
+    with caplog.at_level(logging.WARNING):
+        emitted = warn_if_truncated("length", max_tokens=16, model="gpt-4o-mini")
+
+    assert emitted is True
+    assert len(caplog.records) == 1
+    message = caplog.records[0].getMessage()
+    # The reader needs to know which knob to turn, so the cap has to appear.
+    assert "16" in message
+    assert "gpt-4o-mini" in message
+
+
+def test_silent_on_a_complete_response(caplog):
+    """A complete response must log nothing at all."""
+    with caplog.at_level(logging.WARNING):
+        emitted = warn_if_truncated("stop", max_tokens=16, model="gpt-4o-mini")
+
+    assert emitted is False
+    assert caplog.records == []
+
+
+def test_warns_without_optional_context(caplog):
+    """max_tokens is often unset, and a warning is still better than silence."""
+    with caplog.at_level(logging.WARNING):
+        emitted = warn_if_truncated("max_tokens")
+
+    assert emitted is True
+    assert len(caplog.records) == 1
+
+
+def test_warning_names_whichever_cap_the_provider_actually_sent():
+    """Providers send different cap parameters, so the message must not assume one.
+
+    Gemini sends ``max_output_tokens`` and OpenAI/Azure reasoning models send
+    ``max_completion_tokens``. Naming the wrong setting sends the reader to a
+    knob that will not change anything.
+    """
+    import logging
+
+    from crewai.llms._finish_reason_utils import warn_if_truncated
+
+    for cap, model in [(2048, "gemini-2.5-flash"), (4096, "o3-mini"), (16, "claude-haiku-4-5")]:
+        records = []
+        handler = logging.Handler()
+        handler.emit = records.append  # type: ignore[method-assign]
+        logging.getLogger().addHandler(handler)
+        try:
+            assert warn_if_truncated("length", cap, model) is True
+        finally:
+            logging.getLogger().removeHandler(handler)
+        assert str(cap) in records[0].getMessage()
+        assert model in records[0].getMessage()
+
+
+def test_recognises_the_responses_api_spelling():
+    """The Responses API reports ``max_output_tokens``, not ``length``.
+
+    That path exposes ``status`` rather than a finish reason, so the cause comes
+    from ``incomplete_details.reason``. Missing this spelling meant the warning
+    never fired on the Responses API at all.
+    """
+    from crewai.llms._finish_reason_utils import is_truncated
+
+    assert is_truncated("max_output_tokens") is True
+    assert is_truncated("maxOutputTokens") is True
+    assert is_truncated("content_filter") is False
+
+
+def test_responses_truncation_reason_reads_incomplete_details():
+    """Only an incomplete response yields a reason, and both shapes are accepted."""
+    from types import SimpleNamespace
+
+    from crewai.llms.providers.openai.completion import OpenAICompletion
+
+    read = OpenAICompletion._responses_truncation_reason
+
+    incomplete = SimpleNamespace(
+        status="incomplete", incomplete_details=SimpleNamespace(reason="max_output_tokens")
+    )
+    assert read(incomplete) == "max_output_tokens"
+
+    as_dict = SimpleNamespace(
+        status="incomplete", incomplete_details={"reason": "max_output_tokens"}
+    )
+    assert read(as_dict) == "max_output_tokens"
+
+    assert read(SimpleNamespace(status="completed", incomplete_details=None)) is None
+    assert read(SimpleNamespace(status="incomplete", incomplete_details=None)) is None