diff --git a/python/tests/core/llm/test_llm_retry.py b/python/tests/core/llm/test_llm_retry.py
index a98a7ab0..e095c7f7 100644
--- a/python/tests/core/llm/test_llm_retry.py
+++ b/python/tests/core/llm/test_llm_retry.py
@@ -279,6 +279,103 @@ async def mock_sleep(delay):
 
         assert delays == [2.5]
 
+    @pytest.mark.asyncio
+    async def test_retry_after_capped_at_max_delay(self):
+        """A huge Retry-After must not turn into a multi-minute in-place sleep."""
+        from timbal.core.llm.retry import MAX_RETRY_DELAY
+
+        response = MagicMock()
+        response.headers = {"Retry-After": "999"}
+        delays = []
+
+        async def rate_limited_stream():
+            raise OpenAIRateLimitError("Rate limit exceeded", response=response, body=None)
+            yield
+
+        async def mock_sleep(delay):
+            delays.append(delay)
+
+        with patch("asyncio.sleep", side_effect=mock_sleep):
+            with patch("timbal.core.llm.retry.random.uniform", return_value=0.25):
+                with pytest.raises(OpenAIRateLimitError):
+                    async for _ in _retry_on_error(rate_limited_stream, max_retries=1, retry_delay=1.0, context="Test"):
+                        pass
+
+        assert delays == [MAX_RETRY_DELAY]
+
+    @pytest.mark.asyncio
+    async def test_fail_fast_rate_limit_raises_without_retry(self):
+        """With fail_fast_rate_limit, rate limits raise immediately — no sleep, no retry."""
+        attempt_count = 0
+        response = MagicMock()
+        response.headers = {"Retry-After": "999"}
+        delays = []
+
+        async def rate_limited_stream():
+            nonlocal attempt_count
+            attempt_count += 1
+            raise OpenAIRateLimitError("Rate limit exceeded", response=response, body=None)
+            yield
+
+        async def mock_sleep(delay):
+            delays.append(delay)
+
+        with patch("asyncio.sleep", side_effect=mock_sleep):
+            with pytest.raises(OpenAIRateLimitError):
+                async for _ in _retry_on_error(
+                    rate_limited_stream, max_retries=3, retry_delay=1.0, context="Test", fail_fast_rate_limit=True,
+                ):
+                    pass
+
+        assert attempt_count == 1
+        assert delays == []
+
+    @pytest.mark.asyncio
+    async def test_fail_fast_rate_limit_applies_to_429_status(self):
+        """429 APIStatusError (not just RateLimitError) also fails fast."""
+        attempt_count = 0
+        response = MagicMock()
+        response.status_code = 429
+        response.headers = {}
+
+        async def rate_limited_stream():
+            nonlocal attempt_count
+            attempt_count += 1
+            raise OpenAIAPIStatusError("Too many requests", response=response, body=None)
+            yield
+
+        with pytest.raises(OpenAIAPIStatusError):
+            async for _ in _retry_on_error(
+                rate_limited_stream, max_retries=3, retry_delay=0.01, context="Test", fail_fast_rate_limit=True,
+            ):
+                pass
+
+        assert attempt_count == 1
+
+    @pytest.mark.asyncio
+    async def test_fail_fast_rate_limit_still_retries_server_errors(self):
+        """fail_fast_rate_limit only affects rate limits — 5xx still retries in place."""
+        attempt_count = 0
+        response = MagicMock()
+        response.status_code = 503
+        response.headers = {}
+
+        async def flaky_stream():
+            nonlocal attempt_count
+            attempt_count += 1
+            if attempt_count < 2:
+                raise OpenAIAPIStatusError("Service unavailable", response=response, body=None)
+            yield "success"
+
+        chunks = []
+        async for chunk in _retry_on_error(
+            flaky_stream, max_retries=2, retry_delay=0.01, context="Test", fail_fast_rate_limit=True,
+        ):
+            chunks.append(chunk)
+
+        assert chunks == ["success"]
+        assert attempt_count == 2
+
     @pytest.mark.asyncio
     async def test_invalid_retry_after_uses_jitter_delay(self):
         """Test that malformed Retry-After headers are ignored."""
diff --git a/python/tests/core/test_fallback_model.py b/python/tests/core/test_fallback_model.py
index 7511e931..2401c2f3 100644
--- a/python/tests/core/test_fallback_model.py
+++ b/python/tests/core/test_fallback_model.py
@@ -50,6 +50,28 @@ async def router(**kwargs):
         assert [call["model"] for call in calls] == ["openai/primary", "openai/backup"]
         assert all(call["temperature"] == 0.2 for call in calls)
 
+    @pytest.mark.asyncio
+    async def test_fail_fast_rate_limit_set_for_all_but_last_entry(self):
+        """Every entry with a fallback behind it must fail over on 429 instead of
+        sleeping through Retry-After in place; the last entry retries normally."""
+        model = FallbackModel("openai/primary", "openai/middle", "openai/last")
+        calls = []
+
+        async def router(**kwargs):
+            calls.append(kwargs)
+            if kwargs["model"] != "openai/last":
+                raise _status_error(429)
+            yield "ok"
+
+        chunks = [chunk async for chunk in model.route(router)]
+
+        assert chunks == ["ok"]
+        assert [(call["model"], call["fail_fast_rate_limit"]) for call in calls] == [
+            ("openai/primary", True),
+            ("openai/middle", True),
+            ("openai/last", False),
+        ]
+
     @pytest.mark.asyncio
     async def test_uses_per_entry_retry_and_auth_overrides(self):
         model = FallbackModel(
diff --git a/python/timbal/core/fallback_model.py b/python/timbal/core/fallback_model.py
index 11e3f82c..a8ef6f80 100644
--- a/python/timbal/core/fallback_model.py
+++ b/python/timbal/core/fallback_model.py
@@ -75,11 +75,17 @@ async def route(
 
         for index, entry in enumerate(self.entries):
             started = False
+            has_fallback = index + 1 < len(self.entries)
             kwargs = {
                 **llm_router_kwargs,
                 "model": entry.model,
                 "max_retries": entry.max_retries,
                 "retry_delay": entry.retry_delay,
+                # A rate limit means the provider is unavailable for a while by
+                # definition — while another model is still available, fail over
+                # immediately instead of sleeping through Retry-After in place.
+                # The last entry has nowhere to go, so it retries normally.
+                "fail_fast_rate_limit": has_fallback,
             }
             if entry.api_key is not None:
                 kwargs["api_key"] = entry.api_key
@@ -98,7 +104,7 @@ async def route(
                     raise
 
                 errors.append((entry.model, exc))
-                next_model = self.entries[index + 1].model if index + 1 < len(self.entries) else None
+                next_model = self.entries[index + 1].model if has_fallback else None
                 logger.warning(
                     "Falling back to next LLM model",
                     failed_model=entry.model,
diff --git a/python/timbal/core/llm/retry.py b/python/timbal/core/llm/retry.py
index a0a5a69f..4090d270 100644
--- a/python/timbal/core/llm/retry.py
+++ b/python/timbal/core/llm/retry.py
@@ -12,7 +12,9 @@
 MAX_RETRY_DELAY = 30.0
 
 
-async def _retry_on_error(async_gen_func, max_retries: int, retry_delay: float, context: str):
+async def _retry_on_error(
+    async_gen_func, max_retries: int, retry_delay: float, context: str, fail_fast_rate_limit: bool = False,
+):
     """Helper to retry an async generator function on transient failures.
 
     Retryable errors (using SDK exception types):
@@ -28,11 +30,20 @@ async def _retry_on_error(async_gen_func, max_retries: int, retry_delay: float,
     - Invalid requests (400, 404)
     - Other 4xx client errors
 
+    Retry delays honor a ``Retry-After`` response header as a floor, but are
+    always capped at ``MAX_RETRY_DELAY`` — a provider-side cooldown of minutes
+    must never turn into an in-place sleep of minutes.
+
     Args:
         async_gen_func: Async callable that returns an async generator
         max_retries: Maximum number of retry attempts
         retry_delay: Base delay for exponential backoff
         context: Description for logging (e.g., "Anthropic API")
+        fail_fast_rate_limit: If True, rate-limit errors (RateLimitError / 429)
+            are raised immediately instead of retried in place. Set by
+            FallbackModel for every entry that still has a fallback behind it:
+            a rate-limited provider is unavailable for a while by definition,
+            so the chain should move on rather than sleep.
 
     Yields:
         Items from the async generator
@@ -103,13 +114,21 @@ async def _retry_on_error(async_gen_func, max_retries: int, retry_delay: float,
                 )
                 raise
 
+            if fail_fast_rate_limit and error_type == "rate_limit":
+                logger.warning(
+                    "Rate limited and a fallback model is configured, failing over without retrying",
+                    context=context,
+                    error=error_msg,
+                )
+                raise
+
         # Retry logic for retryable errors
         if attempt < max_retries:
             cap = min(retry_delay * (2**attempt), MAX_RETRY_DELAY)
             delay = random.uniform(0, cap)
             retry_after = _retry_after_seconds(last_error)
             if retry_after is not None:
-                delay = max(delay, retry_after)
+                delay = min(max(delay, retry_after), MAX_RETRY_DELAY)
             logger.warning(
                 "Retryable error from LLM provider, retrying...",
                 context=context,
diff --git a/python/timbal/core/llm/router.py b/python/timbal/core/llm/router.py
index 98528c79..98a1109e 100644
--- a/python/timbal/core/llm/router.py
+++ b/python/timbal/core/llm/router.py
@@ -29,6 +29,7 @@ async def _llm_router(
     api_key: str | SecretStr | None = None,
     max_retries: int = 0,
     retry_delay: float = 1.0,
+    fail_fast_rate_limit: bool = False,
     provider_params: dict[str, Any] | None = None,
 ) -> Message:  # type: ignore[misc]  # Declared as Message for framework schema generation; runtime is an async generator of provider-specific chunks.
     """
@@ -141,5 +142,7 @@ async def _llm_router(
             provider=provider, config=config, **request_kwargs,
         )
 
-    async for res_chunk in _retry_on_error(create_stream, max_retries, retry_delay, context):
+    async for res_chunk in _retry_on_error(
+        create_stream, max_retries, retry_delay, context, fail_fast_rate_limit=fail_fast_rate_limit,
+    ):
         yield res_chunk  # type: ignore[return-type]