diff --git a/language/deepseek-r1/docker/evaluation_requirements.txt b/language/deepseek-r1/docker/evaluation_requirements.txt
index e7b5a84199..132d643bf6 100644
--- a/language/deepseek-r1/docker/evaluation_requirements.txt
+++ b/language/deepseek-r1/docker/evaluation_requirements.txt
@@ -5,6 +5,13 @@
 numpy
 pylatexenc
 
+# LiveCodeBench imports `anthropic.HUMAN_PROMPT` / `AI_PROMPT` at import time
+# (lcb_runner/prompts/test_output_prediction.py). Both were removed in
+# anthropic 1.0.0, and LiveCodeBench only declares `anthropic>=0.42.0`, so an
+# unconstrained install pulls a 1.x SDK and eval_accuracy.py fails before it
+# grades a single sample. Pin below 1.0 until LiveCodeBench drops the import.
+anthropic<1.0
+
 # Install evaluation submodules as editable packages
 -e submodules/LiveCodeBench
 -e submodules/prm800k
\ No newline at end of file