diff --git a/symai/backend/engines/neurosymbolic/anthropic/engine.py b/symai/backend/engines/neurosymbolic/anthropic/engine.py
index 24cfd7aa..d95cc04e 100644
--- a/symai/backend/engines/neurosymbolic/anthropic/engine.py
+++ b/symai/backend/engines/neurosymbolic/anthropic/engine.py
@@ -392,9 +392,24 @@ def _build_thinking_config(self, thinking_arg, model, spec):
 
         thinking_type = thinking_arg.get("type")
         if thinking_type == "disabled":
+            if model == "claude-fable-5":
+                # Fable rejects an explicit {"type": "disabled"} (thinking is
+                # always on); omitting the parameter is the closest valid form.
+                logger.warning(
+                    "Thinking cannot be disabled on claude-fable-5; omitting the thinking parameter."
+                )
+                return None, None
             return {"type": "disabled"}, None
 
-        if model in {"claude-opus-4-8", "claude-opus-4-7"}:
+        # Models where manual thinking (budget_tokens) is removed entirely — any
+        # thinking request is coerced to adaptive (budget_tokens returns 400 there).
+        if model in {
+            "claude-fable-5",
+            "claude-opus-5",
+            "claude-sonnet-5",
+            "claude-opus-4-8",
+            "claude-opus-4-7",
+        }:
             return {"type": "adaptive"}, thinking_arg.get("effort")
 
         if thinking_type == "adaptive":
diff --git a/symai/backend/engines/neurosymbolic/anthropic/models.py b/symai/backend/engines/neurosymbolic/anthropic/models.py
index bf29dc7c..ec123d00 100644
--- a/symai/backend/engines/neurosymbolic/anthropic/models.py
+++ b/symai/backend/engines/neurosymbolic/anthropic/models.py
@@ -52,6 +52,20 @@ class AnthropicModelSpec:
         sampling=False,
         pricing=ModelPricing(input=10.00, output=50.00, cached_input=1.00),
     ),
+    "claude-opus-5": AnthropicModelSpec(
+        context_tokens=LONG_CONTEXT_1M_TOKENS,
+        response_tokens=128_000,
+        reasoning=True,
+        vision=True,
+        # NOTE: thinking is on by default (an omitted `thinking` runs adaptive);
+        # `disabled` is only accepted at effort high or lower. Sampling kwargs
+        # are rejected as deprecated.
+        adaptive_thinking=True,
+        long_context_1m=True,
+        default_long_context_1m=True,
+        sampling=False,
+        pricing=ModelPricing(input=5.00, output=25.00, cached_input=0.50),
+    ),
     "claude-sonnet-5": AnthropicModelSpec(
         context_tokens=LONG_CONTEXT_1M_TOKENS,
         response_tokens=128_000,