diff --git a/tests/test_meter.py b/tests/test_meter.py
index e90fcdf..a4d043d 100644
--- a/tests/test_meter.py
+++ b/tests/test_meter.py
@@ -4097,6 +4097,62 @@ def message(message_id, model, input_tokens, output_tokens):
         self.assertTrue(row["availability"]["cost"])
         self.assertFalse(row["cost_approx"])
         self.assertGreater(row["cost"], 0)
+        self.assertEqual(row["primary_model"], "claude-sonnet-4-6")
+        self.assertEqual(row["models"], ["sonnet-4-6"])
+        self.assertEqual(row["context"]["latest"], 100)
+        self.assertEqual(row["_context_samples"], [100])
+
+    def synthetic_record(self, usage=None):
+        return {
+            "type": "assistant", "timestamp": "2026-07-02T00:01:00.000Z",
+            "message": {
+                "id": "msg-synthetic", "model": "<synthetic>", "content": [],
+                "usage": usage or {
+                    "input_tokens": 0, "cache_creation_input_tokens": 0,
+                    "cache_read_input_tokens": 0, "output_tokens": 0,
+                },
+                "stop_reason": "stop_sequence",
+            },
+        }
+
+    def test_claude_summary_prices_pseudo_model_records_that_report_tokens(self):
+        objs = [
+            self.claude_usage_row(),
+            self.synthetic_record({"input_tokens": 40, "output_tokens": 5}),
+        ]
+
+        row = meter.claude_summary(self.source("claude"), objs)
+
+        self.assertFalse(row["availability"]["cost"])
+        self.assertEqual(row["primary_model"], "<synthetic>")
+
+    def test_claude_recompute_keeps_cost_available_with_synthetic_records(self):
+        records = [{
+            "type": "assistant", "timestamp": "2026-07-02T00:00:00.000Z",
+            "message": {
+                "id": "msg-1", "model": "claude-sonnet-4-6", "content": [],
+                "usage": {"input_tokens": 10, "cache_creation_input_tokens": 5_000,
+                          "cache_read_input_tokens": 5_000, "output_tokens": 10,
+                          "cache_creation": {
+                              "ephemeral_5m_input_tokens": 5_000,
+                              "ephemeral_1h_input_tokens": 0,
+                          }},
+                "stop_reason": "end_turn",
+            },
+        }, self.synthetic_record()]
+        with tempfile.TemporaryDirectory() as tmp:
+            path = Path(tmp) / "session.jsonl"
+            path.write_text("".join(json.dumps(row) + "\n" for row in records))
+            source = {
+                **self.source("claude"), "path": str(path),
+                "session": path.name,
+            }
+            state = meter.recompute_claude(source)
+
+        self.assertTrue(state["availability"]["cost"])
+        self.assertFalse(state["cost_approx"])
+        self.assertGreater(state["total_cost"], 0)
+        self.assertEqual(len(state["executions"]), 1)
 
     def test_unknown_model_keeps_cache_money_unavailable(self):
         record = {
diff --git a/token_meter/runtimes/claude.py b/token_meter/runtimes/claude.py
index b375201..1c4c445 100644
--- a/token_meter/runtimes/claude.py
+++ b/token_meter/runtimes/claude.py
@@ -32,6 +32,12 @@
 
 
 DEFAULT_MODEL = "claude-sonnet-4-6"
+USAGE_TOKEN_FIELDS = (
+    "input_tokens",
+    "output_tokens",
+    "cache_read_input_tokens",
+    "cache_creation_input_tokens",
+)
 MAX_DETAIL_TURNS = 2_000
 MAX_TOOL_EVENTS = 2_000
 ACTIVITY_TAIL_BYTES = 1024 * 1024
@@ -222,6 +228,20 @@ def _cost_coverage_complete(usage, priced):
     )
 
 
+def _unbilled_pseudo_model(model, usage):
+    """Claude Code records locally generated messages as `<synthetic>`.
+
+    They are not model executions, so counting them misreports the session's
+    model identity, collapses latest context to zero, and inflates the execution
+    count. Require both the pseudo-model name shape and absent token evidence so
+    a record that does carry billable tokens is still priced.
+    """
+    name = str(model or "")
+    if not (name.startswith("<") and name.endswith(">")):
+        return False
+    return not any(_safe_int(usage.get(field)) for field in USAGE_TOKEN_FIELDS)
+
+
 def _compact(value, limit=90):
     value = " ".join(str(value or "").split())
     return value[:limit - 1] + "…" if len(value) > limit else value
@@ -1016,6 +1036,8 @@ def recompute_legacy(self, source):
             )
             if not usage:
                 continue
+            if _unbilled_pseudo_model(rec["model"], usage):
+                continue
             input_complete = input_complete and usage["input_available"]
             output_complete = output_complete and usage["output_available"]
             idx = len(series) + 1
@@ -1299,6 +1321,8 @@ def summarize_legacy(self, source, objs=None):
             )
             if not usage:
                 continue
+            if _unbilled_pseudo_model(rec["model"], usage):
+                continue
             input_complete = input_complete and usage["input_available"]
             output_complete = output_complete and usage["output_available"]
             primary_model = rec["model"] or primary_model