diff --git a/scripts/jev-country-calibrate.py b/scripts/jev-country-calibrate.py
index abe57ac..a7b7098 100644
--- a/scripts/jev-country-calibrate.py
+++ b/scripts/jev-country-calibrate.py
@@ -292,7 +292,9 @@ def parse_answer(resp):
         return None
     return {"choice": choice,
             "probabilities": ans.get("probabilities") or {},
-            "confidence": ans.get("confidence")}
+            "confidence": ans.get("confidence"),
+            "usage": resp.get("usage") or {},   # {input_tokens, output_tokens} -- the cost record
+            "model": resp.get("model")}
 
 
 def cmd_judge(args):
@@ -440,11 +442,30 @@ def join(gold, answers):
     return out
 
 
+def usage_totals(answers):
+    """Sum the token usage recorded on each answer -- the run's actual cost, per Jev's usage field."""
+    total, n = {}, 0
+    for a in answers.values():
+        u = a.get("usage") or {}
+        if u:
+            n += 1
+        for k, v in u.items():
+            if isinstance(v, (int, float)):
+                total[k] = total.get(k, 0) + v
+    return total, n
+
+
 def cmd_report(args):
-    joined = join(load_jsonl(args.gold), load_jsonl(args.answers))
+    answers = load_jsonl(args.answers)
+    joined = join(load_jsonl(args.gold), answers)
     n_gold = sum(1 for r in joined if r["gold"])
     print(f"\n{len(joined)} judged  ({n_gold} gold / {len(joined) - n_gold} suspect)\n")
 
+    tot, n_use = usage_totals(answers)
+    if tot:
+        avg = {k: round(v / n_use, 1) for k, v in tot.items()}
+        print(f"TOKEN COST  total {tot}  over {n_use} calls  |  per-account avg {avg}\n")
+
     print("CALIBRATION  (gold location signals only; precision over decided answers)")
     print(f"  {'confidence bin':<16}{'n':>7}{'precision':>11}{'coverage':>10}{'abstain':>9}")
     for b, d in calibration_table(joined).items():
@@ -542,6 +563,16 @@ def _self_test():
     assert crit["nl"].endswith("the Netherlands."), crit["nl"]
     assert "unknown" in crit and len(crit) == len(opts) + 1
 
+    # usage capture: parse_answer carries the cost record; usage_totals sums it across a run
+    pa = parse_answer({"model": "jev-1.13.0", "answers": {"country": {"choice": "nl",
+                       "confidence": 0.97, "probabilities": {"nl": 0.97}}},
+                       "usage": {"input_tokens": 1461, "output_tokens": 339}})
+    assert pa["usage"]["input_tokens"] == 1461 and pa["model"] == "jev-1.13.0", pa
+    tot, n = usage_totals({"a": {"usage": {"input_tokens": 10, "output_tokens": 3}},
+                           "b": {"usage": {"input_tokens": 20, "output_tokens": 7}},
+                           "c": {}})  # no usage -> not counted
+    assert tot == {"input_tokens": 30, "output_tokens": 10} and n == 2, (tot, n)
+
     print("jev-country-calibrate self-test: OK")