fix(claude-usage-report): price cache-writes at their actual TTL

.claude-plugin/marketplace.json

@@ -7,9 +7,6 @@
     {
       "name": "claude-usage-report",
       "source": "./skills/claude-usage-report",
-      "skills": [
-        "."
-      ],
       "description": "Usage/cost reports with self-refreshing pricing and per-account attribution."
     }
   ]

skills/claude-usage-report/SKILL.md → skills/claude-usage-report/skills/claude-usage-report/SKILL.md

@@ -122,15 +122,17 @@ account-recording hook existed; `mixed: a | b` = a session that switched account
 ## Reference
 
 Model prices ($/MTok), records in effect during the period:
-| Model | Effective | Input | Output | Cache-write (5m) | Cache-read | Note |
-|---|---|---:|---:|---:|---:|---|
+| Model | Effective | Input | Output | Cache-write (5m) | Cache-write (1h) | Cache-read | Note |
+|---|---|---:|---:|---:|---:|---:|---|
 {the RATES block from parser output — one row per applicable record; a model with a
 mid-period price change contributes more than one row}
 
 Caveats:
 - Estimate from local transcripts × public list prices, not a bill — for the authoritative
   number use the Anthropic Console usage dashboard for the period.
-- Cache-write assumes the 5-minute TTL (1.25× input); a 1-hour TTL would be 2× input.
+- Cache-write is priced at each message's actual TTL (5-minute = 1.25× input, 1-hour = 2×
+  input), read from the transcript's `cache_creation` breakdown. Legacy transcripts that
+  lack the breakdown fall back to the 5-minute rate.
 - Opus's 1M-context (`[1m]`) runs bill at standard rates; there is no >200K premium tier.
 - Account attribution comes from this plugin's `SessionStart` hook (`hooks/hooks.json` →
   `record_account.sh` → `record_account.py` → `~/.claude/session-accounts.jsonl`) and is

skills/claude-usage-report/prices.json → skills/claude-usage-report/skills/claude-usage-report/prices.json

@@ -1,12 +1,13 @@
 {
-  "_comment": "Anthropic list prices, $/MTok. Fields: in=input, out=output, cw=5-minute cache write (1.25x input), cr=cache read (0.1x input). Each model has one or more records sorted by 'effective' date; the rate applied to a message is the record with the latest effective date <= that message's (UTC) day. This makes a report use the price in effect on the day being reported. Validate against current published pricing when generating a report and append a new dated record when a price changes (see SKILL.md). Known gap: if a price changes between two report runs, days billed before this file was updated will be costed at the older record.",
+  "_comment": "Anthropic list prices, $/MTok. Fields: in=input, out=output, cw=5-minute cache write (1.25x input), cw_1h=1-hour cache write (2x input), cr=cache read (0.1x input). Each message's cache-write tokens are split into 5m/1h from the transcript's cache_creation breakdown and costed at cw / cw_1h respectively; if a record omits cw_1h the parser falls back to 2x in. Each model has one or more records sorted by 'effective' date; the rate applied to a message is the record with the latest effective date <= that message's (UTC) day. This makes a report use the price in effect on the day being reported. Validate against current published pricing when generating a report and append a new dated record when a price changes (see SKILL.md). Known gap: if a price changes between two report runs, days billed before this file was updated will be costed at the older record.",
   "models": {
     "claude-fable-5": [
       {
         "effective": "2000-01-01",
         "in": 10.0,
         "out": 50.0,
         "cw": 12.5,
+        "cw_1h": 20.0,
         "cr": 1.0
       }
     ],
@@ -16,6 +17,7 @@
         "in": 5.0,
         "out": 25.0,
         "cw": 6.25,
+        "cw_1h": 10.0,
         "cr": 0.5
       }
     ],
@@ -25,6 +27,7 @@
         "in": 5.0,
         "out": 25.0,
         "cw": 6.25,
+        "cw_1h": 10.0,
         "cr": 0.5
       }
     ],
@@ -34,6 +37,7 @@
         "in": 5.0,
         "out": 25.0,
         "cw": 6.25,
+        "cw_1h": 10.0,
         "cr": 0.5
       }
     ],
@@ -43,6 +47,7 @@
         "in": 2.0,
         "out": 10.0,
         "cw": 2.5,
+        "cw_1h": 4.0,
         "cr": 0.2,
         "note": "introductory through 2026-08-31"
       },
@@ -51,6 +56,7 @@
         "in": 3.0,
         "out": 15.0,
         "cw": 3.75,
+        "cw_1h": 6.0,
         "cr": 0.3,
         "note": "standard"
       }
@@ -61,6 +67,7 @@
         "in": 3.0,
         "out": 15.0,
         "cw": 3.75,
+        "cw_1h": 6.0,
         "cr": 0.3
       }
     ],
@@ -70,6 +77,7 @@
         "in": 1.0,
         "out": 5.0,
         "cw": 1.25,
+        "cw_1h": 2.0,
         "cr": 0.1
       }
     ],
@@ -145,7 +153,7 @@
     ]
   },
   "_meta": {
-    "last_checked": "2026-07-07T16:38:34",
+    "last_checked": "2026-07-16T19:46:42",
     "source": "https://platform.claude.com/docs/en/about-claude/pricing.md"
   }
 }

skills/claude-usage-report/usage_report.py → skills/claude-usage-report/skills/claude-usage-report/usage_report.py

@@ -62,8 +62,15 @@ def rate_for(model, day):
         else: break
     return best  # None if day precedes the earliest record
 
+def cw1_rate(rec):
+    # 1-hour cache-write rate; fall back to 2x input if a record predates the cw_1h column
+    return rec.get("cw_1h", 2*rec["in"])
+
 def msg_cost(rec, u):
-    return 0.0 if not rec else sum(u[f]*rec[f] for f,_ in FIELDS)/1e6
+    if not rec: return 0.0
+    c = sum(u[f]*rec[f] for f,_ in FIELDS if f != "cw")
+    c += u["cw5"]*rec["cw"] + u["cw1"]*cw1_rate(rec)  # cache-write split by actual TTL
+    return c/1e6
 
 # --- arg parsing: period + optional --top N
 argv = sys.argv[1:]
@@ -157,12 +164,24 @@ def acct_of(sid):
                         if model:
                             usage = msg["usage"]
                             u = {fk: (usage.get(k,0) or 0) for fk,k in FIELDS}
+                            # split cache-write by TTL; legacy transcripts w/o the
+                            # breakdown fall back to the combined field, costed as 5m
+                            cc = usage.get("cache_creation") or {}
+                            if "ephemeral_5m_input_tokens" in cc or "ephemeral_1h_input_tokens" in cc:
+                                u["cw5"] = cc.get("ephemeral_5m_input_tokens",0) or 0
+                                u["cw1"] = cc.get("ephemeral_1h_input_tokens",0) or 0
+                            else:
+                                u["cw5"] = u["cw"]; u["cw1"] = 0
                             rec = rate_for(model, day)
                             if rec is None: unpriced.add(model)
-                            c = msg_cost(rec, u); t = sum(u.values())
+                            c = msg_cost(rec, u); t = sum(u[fk] for fk,_ in FIELDS)
                             for fk,_ in FIELDS:
                                 model_tok[model][fk] += u[fk]
-                                if rec: model_catcost[model][fk] += u[fk]*rec[fk]/1e6
+                                if not rec: continue
+                                if fk == "cw":
+                                    model_catcost[model][fk] += (u["cw5"]*rec["cw"] + u["cw1"]*cw1_rate(rec))/1e6
+                                else:
+                                    model_catcost[model][fk] += u[fk]*rec[fk]/1e6
                             model_cost[model] += c
                             sess_cost[sid] += c; sess_tok[sid] += t
                             day_cost[day] += c; day_tok[day] += t; day_sids[day].add(sid)
@@ -199,13 +218,13 @@ def applicable_records(model):
 if unpriced:
     print(f"WARNING: no price record covers these models for part of the period: {', '.join(sorted(unpriced))} — add/extend them in prices.json.")
 
-print("\nRATES in effect this period ($/MTok): model | effective | in | out | cache-write | cache-read | note")
+print("\nRATES in effect this period ($/MTok): model | effective | in | out | cache-write(5m) | cache-write(1h) | cache-read | note")
 for m in sorted(model_tok, key=lambda m: model_cost[m], reverse=True):
     recs = applicable_records(m)
     if not recs:
         print(f"  {m} | (no price record — UNPRICED)")
     for r in recs:
-        print(f"  {m} | {r['effective']} | {r['in']} | {r['out']} | {r['cw']} | {r['cr']} | {r.get('note','')}")
+        print(f"  {m} | {r['effective']} | {r['in']} | {r['out']} | {r['cw']} | {cw1_rate(r)} | {r['cr']} | {r.get('note','')}")
 
 print("\nBY MODEL: model | total$ | in(tok,$) | out(tok,$) | cache-write(tok,$) | cache-read(tok,$)")
 for m in sorted(model_tok, key=lambda m: model_cost[m], reverse=True):