diff --git a/pyproject.toml b/pyproject.toml
index fc07d7c2e6..503eac4e6d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "free-claude-code"
-version = "6.2.53"
+version = "6.2.54"
 description = "Local proxy connecting coding agents to OpenAI-compatible AI providers"
 readme = "README.md"
 requires-python = ">=3.14.0"
diff --git a/src/free_claude_code/providers/groq/client.py b/src/free_claude_code/providers/groq/client.py
index 1c480190ab..cb83238223 100644
--- a/src/free_claude_code/providers/groq/client.py
+++ b/src/free_claude_code/providers/groq/client.py
@@ -10,6 +10,7 @@
 from loguru import logger
 
 from free_claude_code.core.anthropic import ReasoningReplayMode
+from free_claude_code.core.failures import ExecutionFailure, FailureKind
 from free_claude_code.core.json_types import JsonObject
 from free_claude_code.core.reasoning import (
     ReasoningEffort,
@@ -27,7 +28,7 @@
     validate_extra_body_does_not_override_reasoning_fields,
 )
 
-from .tpm import correct_tpm_completion_budget
+from .tpm import correct_tpm_completion_budget, token_limit_message
 
 _GROQ_EFFORTS = (
     (ReasoningEffort.MINIMAL, "low"),
@@ -102,6 +103,18 @@ def __init__(self) -> None:
         super().__init__(_PROFILE)
         self._model_reasoning_vocabularies: dict[str, frozenset[str]] = {}
 
+    def failure_override(self, error: Exception) -> ExecutionFailure | None:
+        if token_limit_message(error) is None:
+            return None
+        return ExecutionFailure(
+            FailureKind.INVALID_REQUEST,
+            400,
+            "Groq rejected this request because it exceeds the account's token-rate "
+            "allowance. Reduce the request's token budget or use a model/provider "
+            "with a higher allowance.",
+            False,
+        )
+
     def finalize_chat_body(
         self,
         body: dict[str, Any],
diff --git a/src/free_claude_code/providers/groq/tpm.py b/src/free_claude_code/providers/groq/tpm.py
index 78cfc4b937..9bf0316251 100644
--- a/src/free_claude_code/providers/groq/tpm.py
+++ b/src/free_claude_code/providers/groq/tpm.py
@@ -1,4 +1,4 @@
-"""Strict request-local correction for Groq TPM-size rejections."""
+"""Recognize Groq token-limit rejections and correct combined TPM output budgets."""
 
 import re
 from collections.abc import Mapping
@@ -31,11 +31,8 @@ class GroqTpmCorrection:
     corrected_max_completion_tokens: int
 
 
-def correct_tpm_completion_budget(
-    error: Exception,
-    body: Mapping[str, JsonValue],
-) -> GroqTpmCorrection | None:
-    """Subtract Groq's reported TPM overage from this request's output cap."""
+def token_limit_message(error: Exception) -> str | None:
+    """Recognize Groq's structured token-rate rejection, regardless of quota wording."""
     if _status_code(error) != 413:
         return None
 
@@ -48,7 +45,16 @@ def correct_tpm_completion_budget(
         return None
 
     message = detail.get("message")
-    if not isinstance(message, str):
+    return message if isinstance(message, str) else None
+
+
+def correct_tpm_completion_budget(
+    error: Exception,
+    body: Mapping[str, JsonValue],
+) -> GroqTpmCorrection | None:
+    """Subtract Groq's reported TPM overage from this request's output cap."""
+    message = token_limit_message(error)
+    if message is None:
         return None
     if sum(1 for _ in _TPM_MARKER.finditer(message)) != 1:
         return None
diff --git a/tests/api/test_groq_token_limits.py b/tests/api/test_groq_token_limits.py
new file mode 100644
index 0000000000..b3a65ad251
--- /dev/null
+++ b/tests/api/test_groq_token_limits.py
@@ -0,0 +1,67 @@
+import asyncio
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from fastapi.testclient import TestClient
+
+from tests.api.support import create_test_app
+from tests.providers.test_groq_tpm import (
+    _ITPM_MESSAGE,
+    _MODEL,
+    _ORIGINAL_MAX,
+    _detail,
+    _provider,
+    _status_error,
+)
+
+
+@pytest.mark.parametrize(
+    ("wire", "stream"), [("messages", False), ("messages", True), ("responses", True)]
+)
+@pytest.mark.parametrize("quota", ["itpm", "tpm"])
+def test_groq_quota_failure_reaches_clients_after_correction(wire, stream, quota):
+    provider = _provider()
+    error = (
+        _status_error(detail=_detail(_ITPM_MESSAGE))
+        if quota == "itpm"
+        else _status_error()
+    )
+    create = AsyncMock(side_effect=error)
+    payload = {"model": f"groq/{_MODEL}", "stream": stream}
+    if wire == "messages":
+        payload.update(
+            messages=[{"role": "user", "content": "Hello"}], max_tokens=_ORIGINAL_MAX
+        )
+    else:
+        payload.update(input="Hello", max_output_tokens=_ORIGINAL_MAX)
+    try:
+        with (
+            patch(
+                "free_claude_code.api.routes.resolve_provider", return_value=provider
+            ),
+            patch.object(provider._client.chat.completions, "create", create),
+            TestClient(create_test_app()) as client,
+        ):
+            response = client.post(f"/v1/{wire}", json=payload)
+        assert response.status_code == 400
+        assert response.headers["x-should-retry"] == "false"
+        body = response.json()["error"]
+        assert body["type"] == "invalid_request_error"
+        assert "token-rate allowance" in body["message"]
+        assert (
+            "Limit 7000, Requested 23253" in body["message"]
+            if quota == "itpm"
+            else "Limit 8000" in body["message"]
+        )
+        assert response.headers["request-id"] in body["message"]
+        assert create.await_count == (1 if quota == "itpm" else 2)
+        assert (
+            create.await_args_list[0].kwargs["max_completion_tokens"] == _ORIGINAL_MAX
+        )
+        if quota == "tpm":
+            assert (
+                create.await_args_list[1].kwargs["max_completion_tokens"]
+                < _ORIGINAL_MAX
+            )
+    finally:
+        asyncio.run(provider.cleanup())
diff --git a/tests/providers/test_groq_tpm.py b/tests/providers/test_groq_tpm.py
index ae6077c959..b4cefef428 100644
--- a/tests/providers/test_groq_tpm.py
+++ b/tests/providers/test_groq_tpm.py
@@ -9,9 +9,12 @@
 
 from free_claude_code.config.provider_catalog import GROQ_DEFAULT_BASE
 from free_claude_code.core.anthropic.stream_contracts import parse_sse_text
+from free_claude_code.core.failures import FailureKind
 from free_claude_code.core.reasoning import ReasoningEffort, ReasoningPolicy
 from free_claude_code.providers.admission import ProviderOperationKind
+from free_claude_code.providers.failure_policy import classify_provider_failure
 from free_claude_code.providers.groq import GroqProvider
+from free_claude_code.providers.groq.client import GroqChatBehavior
 from free_claude_code.providers.groq.tpm import correct_tpm_completion_budget
 from free_claude_code.providers.request_recovery import RequestRecovery
 from tests.providers.request_factory import make_messages_request
@@ -26,6 +29,29 @@
 _REQUESTED = 26_206
 _ORIGINAL_MAX = 24_576
 _CORRECTED_MAX = 6_370
+_ITPM_MESSAGE = (
+    "Request too large for model `qwen/qwen3.8-27b` on input tokens per minute "
+    "(ITPM): Limit 7000, Requested 23253, please reduce your message size"
+)
+
+
+@pytest.mark.parametrize("wrapped", [False, True])
+@pytest.mark.parametrize("message", [_ITPM_MESSAGE, "Token allowance exceeded"])
+def test_token_limit_failure_is_actionable_without_parsing_prose(wrapped, message):
+    error = _status_error(detail=_detail(message), wrapped=wrapped)
+    failure = classify_provider_failure(
+        error,
+        provider_name="GROQ",
+        read_timeout_s=30,
+        request_id="req_quota",
+        provider_failure_override=GroqChatBehavior().failure_override,
+    )
+    assert failure.status_code == 400
+    assert failure.kind == FailureKind.INVALID_REQUEST
+    assert failure.retryable is False
+    assert message in failure.message
+    assert "req_quota" in failure.message
+    assert "higher" in failure.message
 
 
 def _detail(
@@ -58,6 +84,23 @@ def _status_error(
     )
 
 
+@pytest.mark.parametrize(
+    "error",
+    [
+        _status_error(status=429),
+        _status_error(status=401),
+        _status_error(detail=_detail(error_type="invalid_request_error")),
+        _status_error(detail=_detail(code="request_too_large")),
+        _status_error(detail={"message": "body too large"}),
+        _status_error(detail={**_detail(), "message": 42}),
+        _status_error(detail={**_detail(), "error": _detail()}),
+        _status_error(detail="not JSON"),
+    ],
+)
+def test_unrelated_or_ambiguous_errors_defer_to_shared_policy(error):
+    assert GroqChatBehavior().failure_override(error) is None
+
+
 def _body(max_completion_tokens: object = _ORIGINAL_MAX) -> dict:
     return {
         "model": _MODEL,
@@ -117,6 +160,17 @@ def test_exact_tpm_rejection_corrects_only_completion_budget(wrapped: bool) -> N
     assert body == original
 
 
+@pytest.mark.parametrize(
+    "quota", ["input tokens per minute (ITPM)", "output tokens per minute (OTPM)"]
+)
+def test_separate_token_quotas_do_not_use_combined_tpm_arithmetic(quota):
+    error = _status_error(detail=_detail(f"{quota}: Limit 8000, Requested 26206"))
+    assert correct_tpm_completion_budget(error, _body()) is None
+    failure = GroqChatBehavior().failure_override(error)
+    assert failure is not None
+    assert failure.status_code == 400
+
+
 @pytest.mark.parametrize(
     ("error", "body"),
     [
diff --git a/uv.lock b/uv.lock
index d6b7614d91..60b9b381e5 100644
--- a/uv.lock
+++ b/uv.lock
@@ -599,7 +599,7 @@ wheels = [
 
 [[package]]
 name = "free-claude-code"
-version = "6.2.53"
+version = "6.2.54"
 source = { editable = "." }
 dependencies = [
     { name = "aiohttp" },