diff --git a/openhands-sdk/openhands/sdk/llm/utils/model_features.py b/openhands-sdk/openhands/sdk/llm/utils/model_features.py
index 6321ae78a1..0c3f7abb21 100644
--- a/openhands-sdk/openhands/sdk/llm/utils/model_features.py
+++ b/openhands-sdk/openhands/sdk/llm/utils/model_features.py
@@ -102,6 +102,12 @@ def _normalized_supported_openai_params(model: str | None) -> frozenset[str]:
 REASONING_EFFORT_MODELS: list[str] = [
     # https://www.anthropic.com/news/claude-fable-5
     "claude-fable-5",
+    # LiteLLM recognizes the first-party "anthropic/claude-opus-4-8" id, but not
+    # the Bedrock cross-region inference ids (e.g.
+    # "bedrock/us.anthropic.claude-opus-4-8-v1:0"), which fall through to the
+    # non-reasoning branch and leak temperature/top_p. List explicitly until
+    # LiteLLM ships Bedrock metadata for this model.
+    "claude-opus-4-8",
 ]
 
 
diff --git a/tests/sdk/llm/test_chat_options.py b/tests/sdk/llm/test_chat_options.py
index 4f63e95270..091910ab3f 100644
--- a/tests/sdk/llm/test_chat_options.py
+++ b/tests/sdk/llm/test_chat_options.py
@@ -147,6 +147,32 @@ def test_claude_sonnet_4_6_strips_temp_and_top_p():
     assert "top_p" not in out
 
 
+def test_bedrock_opus_4_8_strips_temp_top_p_without_thinking_block():
+    """Bedrock cross-region claude-opus-4-8 routes through the reasoning path.
+
+    LiteLLM does not (yet) recognize the Bedrock cross-region inference id as a
+    reasoning model, so the SDK-side override must mark it as one. It must take
+    the reasoning_effort path (which strips temperature/top_p) and NOT the
+    extended-thinking path, which would inject the legacy
+    ``thinking.type=enabled`` block + ``interleaved-thinking`` header that
+    Anthropic now rejects for this model (see reverted #3427 / revert #3441).
+    """
+    llm = DummyLLM(
+        model="bedrock/us.anthropic.claude-opus-4-8-v1:0",
+        top_p=1.0,  # SDK default
+        temperature=0.0,  # Often overridden by benchmarks (e.g. SWE-bench)
+        reasoning_effort="high",
+    )
+    out = select_chat_options(llm, user_kwargs={}, has_tools=True)
+
+    assert "temperature" not in out
+    assert "top_p" not in out
+    assert out.get("reasoning_effort") == "high"
+    # Must NOT take the legacy extended-thinking path.
+    assert "thinking" not in out
+    assert "anthropic-beta" not in out.get("extra_headers", {})
+
+
 def test_extended_thinking_budget_clamped_below_max_tokens():
     """Test that thinking.budget_tokens is clamped to max_output_tokens - 1."""
     # Case 1: extended_thinking_budget exceeds max_output_tokens
diff --git a/tests/sdk/llm/test_model_features.py b/tests/sdk/llm/test_model_features.py
index ce69058f53..0aae7e6339 100644
--- a/tests/sdk/llm/test_model_features.py
+++ b/tests/sdk/llm/test_model_features.py
@@ -68,6 +68,16 @@ def test_model_matches(name, pattern, expected):
         ("claude-fable-5", True),
         ("anthropic/claude-fable-5", True),
         ("litellm_proxy/anthropic/claude-fable-5", True),
+        # claude-opus-4-8: LiteLLM recognizes the first-party id, but not the
+        # Bedrock cross-region inference ids, which must be caught by the
+        # SDK-side override so temperature/top_p are stripped before the request
+        # reaches Anthropic (which rejects temperature for this model).
+        ("claude-opus-4-8", True),
+        ("anthropic/claude-opus-4-8", True),
+        ("bedrock/us.anthropic.claude-opus-4-8-v1:0", True),
+        ("bedrock/eu.anthropic.claude-opus-4-8-v1:0", True),
+        ("bedrock/apac.anthropic.claude-opus-4-8-v1:0", True),
+        ("bedrock/global.anthropic.claude-opus-4-8-v1:0", True),
         # LiteLLM proxy with deployment path prefixes (prod/, dev/, staging/, test/)
         ("litellm_proxy/prod/claude-opus-4-5-20251101", True),
         ("litellm_proxy/dev/claude-opus-4-5", True),