diff --git a/.env.example b/.env.example
index 6aff565..db1e1b0 100644
--- a/.env.example
+++ b/.env.example
@@ -34,3 +34,9 @@ GH_TOKEN=
 # PR_AF_NO_BUDGET=false
 # PR_AF_MAX_DURATION_SECONDS=300
 # PR_AF_MAX_COST_USD=2.0
+
+# Concurrency for review_dimension fan-out. Default 10 lets all 6–8
+# dimensions run in parallel; raise/lower based on your provider's
+# per-key rate limits. Was 3 before — caused 3× wall-clock multiplier
+# in production.
+# PR_AF_MAX_CONCURRENT_REVIEWERS=10
diff --git a/src/pr_af/config.py b/src/pr_af/config.py
index 04bdc2d..5215e8b 100644
--- a/src/pr_af/config.py
+++ b/src/pr_af/config.py
@@ -45,9 +45,17 @@ class BudgetConfig(BaseModel):
         }
     )
 
-    # Concurrency — kept low to avoid cascading rate-limit backoff
-    # when using OpenRouter or other rate-limited providers.
-    max_concurrent_reviewers: int = 3
+    # Concurrency for review_dimension fan-out.
+    #
+    # Production data showed 8 review_dimensions throttled by this semaphore at
+    # the previous default of 3, turning per-dimension cost (~25 min) into a
+    # 3× wall-clock multiplier (≥75 min for the review phase alone). Bumped to
+    # 10 — well within OpenRouter's per-key rate limits on Kimi K2.5 and the
+    # other models we run through opencode. Override via PR_AF_MAX_CONCURRENT_REVIEWERS
+    # if a deployment needs to dial it back for a stricter rate-limit ceiling.
+    max_concurrent_reviewers: int = Field(
+        default_factory=lambda: int(os.getenv("PR_AF_MAX_CONCURRENT_REVIEWERS", "10"))
+    )
 
     # Stagger delay (seconds) between launching parallel tasks to avoid
     # burst rate-limit hits.  Set to 0 to disable staggering.
diff --git a/tests/test_staggered_gather.py b/tests/test_staggered_gather.py
index d4e45d8..a67c7ea 100644
--- a/tests/test_staggered_gather.py
+++ b/tests/test_staggered_gather.py
@@ -90,7 +90,21 @@ async def fail() -> str:
 
 
 def test_budget_config_defaults():
-    """Verify the updated concurrency and stagger defaults."""
-    config = BudgetConfig()
-    assert config.max_concurrent_reviewers == 3
-    assert config.stagger_delay_seconds == 2.0
+    """Verify the updated concurrency and stagger defaults.
+
+    Concurrency was raised from 3 → 10 after production data showed 8
+    review_dimensions throttled by the old semaphore, turning ~25-min
+    per-dimension cost into a 3× wall-clock multiplier. The default is
+    overridable via PR_AF_MAX_CONCURRENT_REVIEWERS — clear that env var
+    so the test pins the in-code default rather than the runtime override.
+    """
+    import os
+
+    prior = os.environ.pop("PR_AF_MAX_CONCURRENT_REVIEWERS", None)
+    try:
+        config = BudgetConfig()
+        assert config.max_concurrent_reviewers == 10
+        assert config.stagger_delay_seconds == 2.0
+    finally:
+        if prior is not None:
+            os.environ["PR_AF_MAX_CONCURRENT_REVIEWERS"] = prior