diff --git a/evalharness/model/runner.py b/evalharness/model/runner.py
index b03a225..fef72b3 100644
--- a/evalharness/model/runner.py
+++ b/evalharness/model/runner.py
@@ -443,6 +443,20 @@ async def generate_predictions(
             preds_by_key[keys[i]] = pred
         preds = [preds_by_key[k] for k in keys]
         usages = [p.get('usage', {}) for p in preds]
+        # include RESTORED predictions' usage (they carry it in the ckpt);
+        # previously only fresh generations counted -> restored benches showed 0
+        fresh_keys = {keys[i] for i, _ in pending}
+        for k, p in preds_by_key.items():
+            if k in fresh_keys:
+                continue  # already counted via one()'s total_usage updates
+            u = p.get('usage') or {}
+            if not any(u.get(kk) for kk in ('input_tokens', 'output_tokens')):
+                continue
+            total_usage = total_usage + Usage(
+                input_tokens=int(u.get('input_tokens', 0) or 0),
+                output_tokens=int(u.get('output_tokens', 0) or 0),
+                total_tokens=int(u.get('total_tokens', 0) or 0),
+                latency_s=float(u.get('latency_s', 0) or 0))
         if status_callback and pending:
             status_callback(f'Generation complete: {len(preds)} responses collected')
         return preds, usages, total_usage