diff --git a/api/streaming.py b/api/streaming.py
index 3acf72be..acd6c4eb 100644
--- a/api/streaming.py
+++ b/api/streaming.py
@@ -2892,10 +2892,9 @@ def _run_agent_streaming(
                 input_tokens = getattr(agent, 'session_prompt_tokens', 0) or 0
                 output_tokens = getattr(agent, 'session_completion_tokens', 0) or 0
                 estimated_cost = getattr(agent, 'session_estimated_cost_usd', None)
-                s.input_tokens = (s.input_tokens or 0) + input_tokens
-                s.output_tokens = (s.output_tokens or 0) + output_tokens
-                if estimated_cost:
-                    s.estimated_cost = (s.estimated_cost or 0) + estimated_cost
+                s.input_tokens = input_tokens
+                s.output_tokens = output_tokens
+                s.estimated_cost = estimated_cost
                 # Persist tool-call summaries even when the final message history only
                 # kept bare tool rows and omitted explicit assistant tool_call IDs.
                 tool_calls = _extract_tool_calls_from_messages(
diff --git a/tests/test_issue1857_usage_overwrite.py b/tests/test_issue1857_usage_overwrite.py
new file mode 100644
index 00000000..5e392b56
--- /dev/null
+++ b/tests/test_issue1857_usage_overwrite.py
@@ -0,0 +1,162 @@
+import queue
+import sys
+import types
+from unittest import mock
+
+
+def test_stream_completion_overwrites_session_usage_with_latest_turn(cleanup_test_sessions):
+    """#1857: completed turns must not add prompt tokens to stale session totals."""
+    import api.streaming as streaming
+
+    saved_snapshots = []
+
+    class FakeSession:
+        def __init__(self):
+            self.session_id = "issue1857_usage_overwrite"
+            self.title = "Existing title"
+            self.workspace = "/tmp"
+            self.model = "gpt-5.4"
+            self.model_provider = None
+            self.profile = None
+            self.personality = None
+            self.messages = [
+                {"role": "user", "content": "old"},
+                {"role": "assistant", "content": "old answer"},
+            ]
+            self.context_messages = list(self.messages)
+            self.input_tokens = 9000
+            self.output_tokens = 800
+            self.estimated_cost = 12.34
+            self.tool_calls = []
+            self.gateway_routing = None
+            self.gateway_routing_history = []
+            self.active_stream_id = None
+            self.pending_user_message = None
+            self.pending_attachments = []
+            self.pending_started_at = None
+            self.context_length = 0
+            self.threshold_tokens = 0
+            self.last_prompt_tokens = 0
+            self.llm_title_generated = True
+
+        def save(self, *args, **kwargs):
+            saved_snapshots.append(
+                {
+                    "input_tokens": self.input_tokens,
+                    "output_tokens": self.output_tokens,
+                    "estimated_cost": self.estimated_cost,
+                    "kwargs": kwargs,
+                }
+            )
+
+        def compact(self):
+            return {
+                "session_id": self.session_id,
+                "title": self.title,
+                "workspace": self.workspace,
+                "model": self.model,
+                "created_at": 0,
+                "updated_at": 0,
+                "pinned": False,
+                "archived": False,
+                "project_id": None,
+                "profile": self.profile,
+                "input_tokens": self.input_tokens,
+                "output_tokens": self.output_tokens,
+                "estimated_cost": self.estimated_cost,
+                "personality": self.personality,
+            }
+
+    class UsageAgent:
+        def __init__(
+            self,
+            model=None,
+            provider=None,
+            base_url=None,
+            api_key=None,
+            platform=None,
+            quiet_mode=False,
+            enabled_toolsets=None,
+            fallback_model=None,
+            session_id=None,
+            session_db=None,
+            stream_delta_callback=None,
+            reasoning_callback=None,
+            tool_progress_callback=None,
+            clarify_callback=None,
+        ):
+            self.session_id = session_id
+            self.context_compressor = None
+            self.session_prompt_tokens = 123
+            self.session_completion_tokens = 45
+            self.session_estimated_cost_usd = 0.067
+            self.reasoning_config = None
+            self.ephemeral_system_prompt = None
+            self._last_error = None
+
+        def run_conversation(self, **kwargs):
+            return {
+                "messages": [
+                    {"role": "user", "content": kwargs["persist_user_message"]},
+                    {"role": "assistant", "content": "new answer"},
+                ]
+            }
+
+        def interrupt(self, _message):
+            pass
+
+    fake_session = FakeSession()
+    fake_stream_id = "stream_issue1857_usage_overwrite"
+    fake_queue = queue.Queue()
+    fake_runtime_module = types.ModuleType("hermes_cli.runtime_provider")
+    fake_runtime_module.resolve_runtime_provider = mock.Mock(
+        return_value={
+            "provider": "openai",
+            "base_url": None,
+            "api_key": "sk-test",
+            "api_mode": "chat_completions",
+            "command": None,
+            "args": [],
+            "credential_pool": None,
+        }
+    )
+    fake_hermes_cli = types.ModuleType("hermes_cli")
+    fake_hermes_cli.runtime_provider = fake_runtime_module
+    fake_hermes_state = types.ModuleType("hermes_state")
+    fake_hermes_state.SessionDB = mock.Mock(return_value=None)
+
+    with mock.patch.object(streaming, "get_session", return_value=fake_session), \
+         mock.patch.object(streaming, "_get_ai_agent", return_value=UsageAgent), \
+         mock.patch.object(streaming, "resolve_model_provider", return_value=("gpt-5.4", "openai", None)), \
+         mock.patch("api.config.get_config", return_value={}), \
+         mock.patch("api.config._resolve_cli_toolsets", return_value=[]), \
+         mock.patch.dict(
+             sys.modules,
+             {
+                 "hermes_cli": fake_hermes_cli,
+                 "hermes_cli.runtime_provider": fake_runtime_module,
+                 "hermes_state": fake_hermes_state,
+             },
+         ):
+        streaming.STREAMS[fake_stream_id] = fake_queue
+        streaming._run_agent_streaming(
+            session_id=fake_session.session_id,
+            msg_text="new turn",
+            model="gpt-5.4",
+            workspace="/tmp",
+            stream_id=fake_stream_id,
+        )
+
+    assert fake_session.input_tokens == 123
+    assert fake_session.output_tokens == 45
+    assert fake_session.estimated_cost == 0.067
+    assert any(
+        event == "done"
+        and payload["usage"]["input_tokens"] == 123
+        and payload["usage"]["output_tokens"] == 45
+        and payload["usage"]["estimated_cost"] == 0.067
+        for event, payload in list(fake_queue.queue)
+    )
+    assert saved_snapshots[-1]["input_tokens"] == 123
+    assert saved_snapshots[-1]["output_tokens"] == 45
+    assert saved_snapshots[-1]["estimated_cost"] == 0.067