diff --git a/backend/apps/agents/agent_manager.py b/backend/apps/agents/agent_manager.py
index 409efe46..4e85e4ec 100644
--- a/backend/apps/agents/agent_manager.py
+++ b/backend/apps/agents/agent_manager.py
@@ -244,6 +244,7 @@ class AgentManager(SessionLifecycle, SessionHistory, SessionPersistence, Messagi
             return
 
         session.status = "running"
+        session.last_failure_kind = None
 
         # Resolve the model id now so every closure (approval hook, tool executed handler, etc.) has both the short name and the 9Router-prefixed id available without re-resolving. The short name is what the user sees; the router id is what 9Router reports its per-model counters under.
         from backend.apps.agents.providers.registry import (
diff --git a/backend/apps/agents/core/models.py b/backend/apps/agents/core/models.py
index 99758bc2..b436f835 100644
--- a/backend/apps/agents/core/models.py
+++ b/backend/apps/agents/core/models.py
@@ -169,6 +169,8 @@ class AgentSession(BaseModel):
     # moment they send again. Failing over is for finishing THE CURRENT ASK; leaving the chat on a
     # metered key forever would bill them per token with one line said about it, once (ENG-383).
     lane_failover_from: Optional[str] = None
+    # How the last turn ended when it ended on a card; a parent reads it off a child so a declined task is not re-dispatched.
+    last_failure_kind: Optional[Literal["policy_block"]] = None
     # Consecutive dirty deaths this session was MID-TURN for; the crash auto-resume breaker (hermes #30719 pairing: auto-resume must never outrun its circuit breaker).
     crash_interrupt_count: int = 0
     # Outage rounds spent on this ask: the in-turn ladder covers only 335s, and the work is checkpointed, so a longer drop is waited out rather than ending the task.
diff --git a/backend/apps/agents/manager/run/TurnRunner.py b/backend/apps/agents/manager/run/TurnRunner.py
index 5cd39465..63b9a680 100644
--- a/backend/apps/agents/manager/run/TurnRunner.py
+++ b/backend/apps/agents/manager/run/TurnRunner.py
@@ -50,7 +50,8 @@ class TurnRunner(AgentManagerProtocol):
         # Deliberate faults, so the guards below get drilled instead of waited for. Inert unless
         # OSW_FAULT names them; a shipped build never sets it. Raised HERE because this is the same
         # door a real provider failure comes through, so the drill exercises the real recovery.
-        if p_fault_armed("policy_block"):
+        # Budgeted like auth_401 (OSW_FAULT_FIRES=N), so a drill can block ONE turn and watch the failover finish the ask.
+        if p_fault_fire("policy_block"):
             # The provider's real wording, so the drill hits the same classifier a field block does.
             raise RuntimeError(
                 "API Error: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":"
diff --git a/backend/apps/agents/manager/run/handle_run_error.py b/backend/apps/agents/manager/run/handle_run_error.py
index e722e9c1..fa3bb842 100644
--- a/backend/apps/agents/manager/run/handle_run_error.py
+++ b/backend/apps/agents/manager/run/handle_run_error.py
@@ -47,7 +47,7 @@ from backend.apps.agents.core.extract_reset_hint import extract_reset_hint
 from backend.apps.agents.core.redact_for_telemetry import redact_for_telemetry
 from backend.apps.agents.core import flight_recorder
 from backend.apps.agents.manager.run.empty_finish import count_tool_calls
-from backend.apps.agents.session_credential import api_key_twin_model
+from backend.apps.agents.session_credential import api_key_twin_model, policy_block_sibling
 
 logger = logging.getLogger(__name__)
 
@@ -502,8 +502,35 @@ async def handle_run_error(e: Exception, session: AgentSession, session_id: str,
                 "this turn, rely on the visible conversation.")
             logger.warning(f"Agent {session_id}: provider content-policy block on a turn carrying a {p_sent} history prefix; retrying with {session.history_prefix_mode}")
             return
+        # Nothing left to strip. The classifier declines Opus 5 where it passes Opus 4.8 (fleet, 2026-09-07:
+        # 78 blocks vs 0 across a comparable error volume), so the step finishes on the same lane's sibling
+        # first: same subscription, same wallet, nothing billed. One ask, announced, and the next real
+        # message puts the chat back on the model the user chose (Messaging restores lane_failover_from).
+        p_sibling = policy_block_sibling(session.model or "") if not session.lane_failover_from else None
+        if p_sibling:
+            p_from = session.model
+            session.model = p_sibling
+            session.lane_failover_from = p_from
+            session.needs_fresh_session = True
+            session.pending_continuation = True
+            session.pending_continuation_prompt = "Continue where you left off and finish the task, then answer in plain text."
+            p_notice = Message(
+                role="system",
+                content=("Anthropic's filter declined this request on Claude Opus 5, which it does far more often "
+                         "than on other models. This step is finishing on Claude Opus 4.8 on the same subscription; "
+                         "your next message goes back to Opus 5. If it keeps happening, switch this chat to Opus 4.8 or Sonnet 5."),
+                branch_id=session.active_branch_id,
+            )
+            absorb_repeat_card(session, p_notice)
+            await ws_manager.send_to_session(session_id, "agent:message", {
+                "session_id": session_id, "message": p_notice.model_dump(mode="json"),
+            })
+            flight_recorder.record_recovery(session_id, "policy_sibling", session.model, 1)
+            logger.warning(f"Agent {session_id}: policy block on {p_from}; finishing this ask on the same-lane sibling {p_sibling}")
+            return
         # The subscription lane declined and nothing is left to strip; fleet data says the same request passes on an API key (0 of 328 vs 4.4%), so a user who connected their own Anthropic key continues there, told in one line, instead of losing the ask (ENG-383).
-        p_twin = api_key_twin_model(session.model or "", load_settings())
+        # One failover per ask, whichever door it took: a borrowed sibling that is blocked again ends on the card, never on a second borrow that bills the key.
+        p_twin = api_key_twin_model(session.model or "", load_settings()) if not session.lane_failover_from else None
         if p_twin:
             p_from = session.model
             session.model = p_twin
@@ -526,6 +553,7 @@ async def handle_run_error(e: Exception, session: AgentSession, session_id: str,
             logger.warning(f"Agent {session_id}: policy block on the subscription lane; failing over {p_from} -> {p_twin}")
             return
         p_retried = session.history_prefix_mode == "none"
+        session.last_failure_kind = "policy_block"
         friendly_msg = (
             "The model provider declined this request (its automated policy filter flagged the "
             "conversation's content)"
diff --git a/backend/apps/agents/session_credential.py b/backend/apps/agents/session_credential.py
index e482f43c..2e7b8a06 100644
--- a/backend/apps/agents/session_credential.py
+++ b/backend/apps/agents/session_credential.py
@@ -189,6 +189,22 @@ def write_would_suicide(field: str, new_value: Any, powering: PoweringCredential
     return False
 
 
+# Anthropic's abuse classifier declines Opus 5 requests it passes on every other model: 78 of 84 fleet
+# blocks in the fortnight to 2026-09-07 were on opus-5 against 0 on opus-4-8 across a comparable error
+# volume, and Anthropic's own block text says to switch models. The sibling is the same vendor on the
+# SAME lane and wallet, so a step can finish there without spending anything the user did not connect.
+POLICY_BLOCK_SIBLING: dict[str, str] = {
+    "opus-5": "opus-4-8",
+    "opus-5-cc": "opus-4-8-cc",
+    "opus-5-api": "opus-4-8-api",
+}
+
+
+def policy_block_sibling(model_value: str) -> Optional[str]:
+    """The same-lane model a policy-blocked step may finish on, or None when there is no such sibling."""
+    return POLICY_BLOCK_SIBLING.get(model_value)
+
+
 def api_key_twin_model(model_value: str, settings: AppSettings) -> Optional[str]:
     """The same Claude model on the user's OWN Anthropic API key, when this run was on a subscription
     lane and such a key is configured; None otherwise. Never the OpenSwarm Pro pool (not their money to
diff --git a/backend/tests/test_content_policy_block.py b/backend/tests/test_content_policy_block.py
index a398122e..70af6a89 100644
--- a/backend/tests/test_content_policy_block.py
+++ b/backend/tests/test_content_policy_block.py
@@ -82,6 +82,7 @@ def test_block_on_a_recap_bearing_turn_drops_to_none_and_retries_silently(monkey
 def test_block_with_no_recap_renders_the_honest_terminal_card(monkeypatch):
     captured: list = []
     s = p_session_with_history()
+    s.model = "sonnet-5"
     s.history_prefix_mode = "none"
     s.history_prefix_sent = "none"
     p_block(s, captured, monkeypatch, text="The agent runtime reported this turn failed (error_during_execution). API Error: 400 https://www.anthrop" + FIELD_TAIL)
@@ -134,11 +135,11 @@ def test_api_key_twin_only_for_a_subscription_lane_with_the_users_own_key():
 def test_block_with_no_recap_fails_over_to_the_users_own_api_key(monkeypatch):
     captured: list = []
     s = p_session_with_history()
-    s.model = "opus-5-cc"
+    s.model = "sonnet-5-cc"
     s.history_prefix_mode = "none"
     s.history_prefix_sent = "none"
     p_block(s, captured, monkeypatch, settings=p_settings(anthropic_api_key="sk-ant-x"))
-    assert s.model == "opus-5-api"
+    assert s.model == "sonnet-5-api"
     assert s.pending_continuation is True and s.needs_fresh_session is True
     cards = [m for m in s.messages if m.role == "system"]
     assert len(cards) == 1 and "API key" in str(cards[0].content) and "declined" not in str(cards[0].content).lower().replace("declined this request on your subscription", "")
@@ -150,9 +151,58 @@ def test_block_with_no_recap_fails_over_to_the_users_own_api_key(monkeypatch):
 def test_block_with_no_recap_and_no_key_still_ends_with_the_card(monkeypatch):
     captured: list = []
     s = p_session_with_history()
-    s.model = "opus-5-cc"
+    s.model = "sonnet-5-cc"
     s.history_prefix_mode = "none"
     s.history_prefix_sent = "none"
     p_block(s, captured, monkeypatch, settings=p_settings())
-    assert s.model == "opus-5-cc" and s.pending_continuation is False
+    assert s.model == "sonnet-5-cc" and s.pending_continuation is False
     assert [m for m in s.messages if m.role == "system"][0].content.startswith("The model provider declined")
+
+
+# Fleet, 14 days to 2026-09-07: 78 policy blocks on opus-5 against 0 on opus-4-8 (87 errors of every other
+# kind there), every one at spawn on the subscription lane. The same lane's sibling is free and passes.
+def test_policy_block_sibling_is_opus_5_to_opus_4_8_on_the_same_lane_only():
+    from backend.apps.agents.session_credential import policy_block_sibling
+    assert policy_block_sibling("opus-5") == "opus-4-8"
+    assert policy_block_sibling("opus-5-cc") == "opus-4-8-cc"
+    assert policy_block_sibling("opus-5-api") == "opus-4-8-api"
+    assert policy_block_sibling("opus-4-8") is None, "the sibling has no sibling, so a block there cards"
+    assert policy_block_sibling("sonnet-5") is None
+    assert policy_block_sibling("gpt-5.6") is None
+
+
+def test_opus_5_block_with_nothing_to_strip_finishes_on_opus_4_8_and_says_so(monkeypatch):
+    captured: list = []
+    s = p_session_with_history()
+    s.model = "opus-5-cc"
+    s.history_prefix_mode = "none"
+    s.history_prefix_sent = "none"
+    p_block(s, captured, monkeypatch, settings=p_settings(anthropic_api_key="sk-ant-x"))
+    assert s.model == "opus-4-8-cc", "the free same-lane sibling comes before the billed API-key twin"
+    assert s.lane_failover_from == "opus-5-cc"
+    assert s.pending_continuation is True and s.needs_fresh_session is True
+    cards = [m for m in s.messages if m.role == "system"]
+    assert len(cards) == 1 and "Opus 4.8" in str(cards[0].content) and "next message goes back" in str(cards[0].content)
+    kinds = [(d.get("kind"), d.get("subkind")) for d in captured]
+    assert ("model_error", "policy_block:none") in kinds, "the block itself is still reported"
+    assert ("recovered", "policy_sibling") in kinds
+
+
+def test_a_block_on_the_sibling_never_chains_to_a_third_model(monkeypatch):
+    """A borrowed ask that is blocked again ends on the card; without this the failover could walk twin -> sibling -> twin."""
+    captured: list = []
+    s = p_session_with_history()
+    s.model = "opus-4-8-cc"
+    s.lane_failover_from = "opus-5-cc"
+    s.history_prefix_mode = "none"
+    s.history_prefix_sent = "none"
+    p_block(s, captured, monkeypatch, settings=p_settings(anthropic_api_key="sk-ant-x"))
+    assert s.model == "opus-4-8-cc" and s.pending_continuation is False
+    assert [m for m in s.messages if m.role == "system"][0].content.startswith("The model provider declined")
+
+
+def test_the_sibling_door_sits_above_the_api_key_twin_door():
+    """Ordering is the contract: the free lane sibling is tried before the user's key is spent."""
+    import inspect
+    src = inspect.getsource(hre.handle_run_error)
+    assert src.index("policy_block_sibling(") < src.index("api_key_twin_model(")