diff --git a/src/scripto/gui/viewmodel.py b/src/scripto/gui/viewmodel.py
index d09bce5..ce10b57 100644
--- a/src/scripto/gui/viewmodel.py
+++ b/src/scripto/gui/viewmodel.py
@@ -120,9 +120,8 @@ def __init__(
 
         self.log_lines: list[str] = []
 
-        # ETA bookkeeping: wall time of completed files this batch
-        self._durations: list[float] = []
-        self._active_since: float | None = None
+        # ETA bookkeeping: when this batch started (see _snapshot)
+        self._batch_started: float | None = None
 
         # History-translation queue: one worker, jobs survive any dialog.
         self.translation_jobs: list[TranslationJob] = []
@@ -185,8 +184,7 @@ def start_batch(
             self._stop.clear()
             self._final_stats = None
             self._finished_flag = False
-            self._durations = []
-            self._active_since = None
+            self._batch_started = time.monotonic()
             self._id_map = {n: ids[n - 1] for n in range(1, len(ids) + 1)}
 
         config = self.config.load()
@@ -260,7 +258,6 @@ def drain(self) -> DrainResult:
                 row = self._row_for_job(int(event.subject.split(":", 1)[1]))
                 if row is None:
                     continue
-                self._track_eta(row, event.status)
                 row.status = event.status
                 row.error = event.detail
                 row.error_key = event.detail_key
@@ -298,14 +295,6 @@ def _row_for_job(self, job_id: int) -> FileRow | None:
         mapped = getattr(self, "_id_map", {}).get(job_id, job_id)
         return self.rows.get(mapped)
 
-    def _track_eta(self, row: FileRow, new_status: str) -> None:
-        now = time.monotonic()
-        if new_status == JobStatus.TRANSCRIBING.value:
-            self._active_since = now
-        elif new_status == JobStatus.DONE.value and self._active_since is not None:
-            self._durations.append(now - self._active_since)
-            self._active_since = None
-
     def _snapshot(self) -> Snapshot:
         rows = list(self.rows.values())
         terminal = {
@@ -315,15 +304,25 @@ def _snapshot(self) -> Snapshot:
         done = sum(1 for r in rows if r.status in terminal)
         active = next(
             (r for r in rows if r.status in (
-                JobStatus.TRANSCRIBING.value, JobStatus.EXTRACTING.value,
-                JobStatus.TRANSLATING.value,
+                JobStatus.DOWNLOADING.value, JobStatus.TRANSCRIBING.value,
+                JobStatus.EXTRACTING.value, JobStatus.TRANSLATING.value,
             )),
             None,
         )
+        # Wall-clock throughput, not the average of per-file transcribe times.
+        # Timing files individually measured only the stage we happened to
+        # watch and charged every remaining file a full transcription — with
+        # a half-subtitled library, where most files are skipped in
+        # milliseconds, that produced estimates in the thousands of minutes.
+        # Elapsed-over-finished counts everything the batch actually spends
+        # (iCloud downloads, extraction stalls, the one-time model load,
+        # translation running alongside) and lets cheap files pull the
+        # average down by themselves.
         eta = None
-        if self.running and self._durations:
+        if self.running and done and self._batch_started is not None:
             remaining = len(rows) - done
-            eta = sum(self._durations) / len(self._durations) * remaining
+            elapsed = time.monotonic() - self._batch_started
+            eta = elapsed / done * remaining
         return Snapshot(
             running=self.running,
             done=done,
diff --git a/tests/test_viewmodel.py b/tests/test_viewmodel.py
index 572cfa2..c3a987e 100644
--- a/tests/test_viewmodel.py
+++ b/tests/test_viewmodel.py
@@ -55,21 +55,48 @@ def test_drain_applies_status_and_progress(tmp_path):
     assert any("boom" in line for line in result.log_lines)
 
 
-def test_eta_from_completed_durations(tmp_path):
-    vm = make_vm(tmp_path)
-    make_media(tmp_path, 3)
+def _running_batch(vm, tmp_path, count: int, *, elapsed: float):
+    make_media(tmp_path, count)
     vm.scan_inputs(str(tmp_path))
     vm.running = True
+    vm._batch_started = time.monotonic() - elapsed
     vm._id_map = {i: rid for i, rid in enumerate(vm.row_order, start=1)}
 
-    vm.bus.emit(StatusEvent(subject="job:1", status="transcribing"))
-    vm.drain()
-    time.sleep(0.05)
+
+def test_eta_extrapolates_wall_clock_throughput(tmp_path):
+    vm = make_vm(tmp_path)
+    _running_batch(vm, tmp_path, 4, elapsed=60.0)
+
     vm.bus.emit(StatusEvent(subject="job:1", status="done"))
-    result = vm.drain()
-    snap = result.snapshot
-    assert snap.done == 1 and snap.total == 3
-    assert snap.eta_sec is not None and snap.eta_sec > 0
+    snap = vm.drain().snapshot
+    assert snap.done == 1 and snap.total == 4
+    assert 170 < snap.eta_sec < 195   # a minute bought one file, three to go
+
+
+def test_eta_treats_instantly_skipped_files_as_cheap(tmp_path):
+    """The shape that produced a 3000-minute estimate: most files skip.
+
+    Timing transcriptions individually and multiplying by the files left
+    charged every already-subtitled file a full transcription. Wall-clock
+    throughput lets them cost what they actually cost.
+    """
+    vm = make_vm(tmp_path)
+    _running_batch(vm, tmp_path, 10, elapsed=5.0)
+
+    for job_id in range(1, 6):   # five files already had subtitles
+        vm.bus.emit(StatusEvent(subject=f"job:{job_id}", status="skipped"))
+    snap = vm.drain().snapshot
+    assert snap.done == 5
+    assert snap.eta_sec < 10     # seconds, not five full transcriptions
+
+
+def test_eta_waits_for_the_first_finished_file(tmp_path):
+    vm = make_vm(tmp_path)
+    _running_batch(vm, tmp_path, 3, elapsed=30.0)
+    vm.bus.emit(StatusEvent(subject="job:1", status="downloading"))
+    snap = vm.drain().snapshot
+    assert snap.eta_sec is None            # nothing to extrapolate from yet
+    assert snap.current_status == "downloading"   # but the row is shown as busy
 
 
 def test_batch_runs_via_fake_pipeline(tmp_path, monkeypatch):