diff --git a/src/helm/benchmark/metrics/gpqa_chain_of_thought_metric.py b/src/helm/benchmark/metrics/gpqa_chain_of_thought_metric.py
index f6d994943b..ea6418c514 100644
--- a/src/helm/benchmark/metrics/gpqa_chain_of_thought_metric.py
+++ b/src/helm/benchmark/metrics/gpqa_chain_of_thought_metric.py
@@ -25,8 +25,8 @@ def extract_answer(output_text: str) -> Optional[str]:
     if match:
         return match.group(1)
 
-    # Second regex: Matches "[answer: (A-J)]" with optional leading characters like "."
-    match = re.search(r"\.*\[aA\]nswer:\s*\(?([A-J])\)?", output_text)
+    # Second regex: Matches "answer: (A-J)" (e.g. "Answer: B" or "[answer: B]") with optional leading characters
+    match = re.search(r"\.*[aA]nswer:\s*\(?([A-J])\)?", output_text)
     if match:
         return match.group(1)
 
diff --git a/src/helm/benchmark/metrics/test_gpqa_chain_of_thought_metric.py b/src/helm/benchmark/metrics/test_gpqa_chain_of_thought_metric.py
new file mode 100644
index 0000000000..fb4cfd6ed8
--- /dev/null
+++ b/src/helm/benchmark/metrics/test_gpqa_chain_of_thought_metric.py
@@ -0,0 +1,21 @@
+from helm.benchmark.metrics.gpqa_chain_of_thought_metric import extract_answer
+
+
+def test_extract_answer_correct_answer_is_format():
+    # Primary format requested by the GPQA chain-of-thought prompt.
+    assert extract_answer("The correct answer is (A)") == "A"
+    assert extract_answer("The correct answer is B") == "B"
+
+
+def test_extract_answer_answer_colon_format():
+    # Fallback format: "Answer: X" / "answer: X" (previously broken by the
+    # literal-escaped brackets in the regex, which only matched "[aA]nswer:").
+    assert extract_answer("Answer: C") == "C"
+    assert extract_answer("The answer: B") == "B"
+    assert extract_answer("answer: E") == "E"
+    assert extract_answer("...Answer: D") == "D"
+    assert extract_answer("[answer: F]") == "F"
+
+
+def test_extract_answer_no_match():
+    assert extract_answer("I am not sure.") is None