diff --git a/studio/backend/core/inference/context_compaction.py b/studio/backend/core/inference/context_compaction.py
index 8428d0cd53..18c36bb07c 100644
--- a/studio/backend/core/inference/context_compaction.py
+++ b/studio/backend/core/inference/context_compaction.py
@@ -65,7 +65,12 @@ def estimate_tokens(messages: list[dict]) -> int:
                 args = (tc.get("function") or {}).get("arguments")
                 if isinstance(args, str):
                     total_chars += len(args)
-    return total_chars // _CHARS_PER_TOKEN
+    # Ceil-style division: floor (`// 4`) would systematically
+    # underestimate non-multiple-of-4 lengths, letting just-over-budget
+    # prompts appear under threshold and bypass compaction — the exact
+    # failure this module exists to prevent. Round up so the heuristic
+    # stays on the conservative side described above.
+    return -(-total_chars // _CHARS_PER_TOKEN)
 
 
 def _is_multimodal(msg: dict) -> bool:
diff --git a/studio/backend/tests/test_context_compaction.py b/studio/backend/tests/test_context_compaction.py
index 03e688e78a..c42b39d9ad 100644
--- a/studio/backend/tests/test_context_compaction.py
+++ b/studio/backend/tests/test_context_compaction.py
@@ -47,6 +47,17 @@ class TestEstimateTokens:
         msgs = [_msg("user", "abcd")]
         assert estimate_tokens(msgs) == 1
 
+    def test_non_multiple_of_4_rounds_up(self):
+        # Regression: floor (`// 4`) would estimate 1 for "abcde" and
+        # let an over-budget message slip past the threshold check.
+        # Ceil keeps the heuristic conservative.
+        msgs = [_msg("user", "abcde")]  # 5 chars
+        assert estimate_tokens(msgs) == 2
+
+    def test_single_char_rounds_up_to_one_token(self):
+        msgs = [_msg("user", "a")]
+        assert estimate_tokens(msgs) == 1
+
     def test_multimodal_text_part_counts(self):
         msgs = [
             {