Merge fix/cache-ttl-pricing: v0.1.2

Author: cballou

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "session-cost",
   "displayName": "Session Cost Tracker",
-  "version": "0.1.1",
+  "version": "0.1.2",
   "description": "Track estimated per-session Claude Code spend, report it with /session-cost:report, and optionally enforce soft/hard budget limits from ~/.claude/budgets.json",
   "author": {
     "name": "OpenSea",

CHANGELOG.md

@@ -25,3 +25,10 @@ Initial release.
 - Fix: remove duplicate `hooks` declaration from plugin.json — `hooks/hooks.json`
   is auto-loaded by convention and redeclaring it fails plugin load with
   "Duplicate hooks file detected". Caught by live install verification.
+
+## 0.1.2 — 2026-06-10
+
+- Fix: price 1-hour-TTL cache writes at 2x base input (5-minute at 1.25x)
+  using the per-TTL `cache_creation` breakdown in transcript usage blocks.
+  Claude Code uses 1h cache exclusively, so flat 1.25x underpriced cache
+  writes by 1.6x. Verified against live pricing docs and a real transcript.

scripts/cost_tracker.py

@@ -24,9 +24,10 @@
     "default":           {"input":  5.00, "output": 25.00},
 }
 
-# Cache multipliers (5-minute TTL writes)
-CACHE_READ_DISCOUNT = 0.10     # cache reads cost ~10% of input price
-CACHE_CREATION_PREMIUM = 1.25  # cache writes cost ~125% of input price
+# Cache multipliers (relative to base input price)
+CACHE_READ_DISCOUNT = 0.10        # cache reads cost 10% of input price
+CACHE_CREATION_PREMIUM = 1.25     # 5-minute-TTL cache writes
+CACHE_CREATION_PREMIUM_1H = 2.0   # 1-hour-TTL cache writes (what Claude Code uses)
 
 # Average token estimates per tool type (input_tokens, output_tokens).
 # Rough averages — the value is in aggregation, not per-call precision.
@@ -90,13 +91,29 @@ def get_pricing(model):
 
 
 def compute_cost(usage, model):
-    """Estimated USD cost from a Messages API usage dict."""
+    """Estimated USD cost from a Messages API usage dict.
+
+    When the usage carries a per-TTL `cache_creation` breakdown (real Claude
+    Code transcripts do, and use 1-hour cache exclusively), price 5m writes
+    at 1.25x and 1h writes at 2x base input. Otherwise fall back to 1.25x on
+    the `cache_creation_input_tokens` total.
+    """
     pricing = get_pricing(model)
+    breakdown = usage.get("cache_creation")
+    if isinstance(breakdown, dict):
+        cache_write_cost = pricing["input"] * (
+            breakdown.get("ephemeral_5m_input_tokens", 0) * CACHE_CREATION_PREMIUM
+            + breakdown.get("ephemeral_1h_input_tokens", 0) * CACHE_CREATION_PREMIUM_1H
+        )
+    else:
+        cache_write_cost = (
+            usage.get("cache_creation_input_tokens", 0)
+            * pricing["input"] * CACHE_CREATION_PREMIUM
+        )
     return (
         usage.get("input_tokens", 0) * pricing["input"]
         + usage.get("output_tokens", 0) * pricing["output"]
-        + usage.get("cache_creation_input_tokens", 0)
-        * pricing["input"] * CACHE_CREATION_PREMIUM
+        + cache_write_cost
         + usage.get("cache_read_input_tokens", 0)
         * pricing["input"] * CACHE_READ_DISCOUNT
     ) / 1_000_000

tests/test_cost_tracker.py

@@ -43,7 +43,7 @@ def test_basic_io_tokens(self):
         cost = cost_tracker.compute_cost(usage, "claude-opus-4-8")
         assert cost == 5.00 + 25.00
 
-    def test_cache_multipliers(self):
+    def test_cache_multipliers_without_ttl_breakdown(self):
         usage = {
             "input_tokens": 0,
             "output_tokens": 0,
@@ -53,6 +53,43 @@ def test_cache_multipliers(self):
         cost = cost_tracker.compute_cost(usage, "claude-opus-4-8")
         assert cost == 5.00 * 1.25 + 5.00 * 0.10
 
+    def test_cache_ttl_breakdown_prices_1h_writes_at_2x(self):
+        """Real Claude Code transcripts carry a cache_creation breakdown and
+        use 1-hour cache exclusively — 1h writes cost 2x base, not 1.25x."""
+        usage = {
+            "cache_creation_input_tokens": 1_500_000,
+            "cache_creation": {
+                "ephemeral_5m_input_tokens": 500_000,
+                "ephemeral_1h_input_tokens": 1_000_000,
+            },
+        }
+        cost = cost_tracker.compute_cost(usage, "claude-opus-4-8")
+        # breakdown replaces the flat 1.25x on the total — no double count
+        assert cost == 0.5 * 5.00 * 1.25 + 1.0 * 5.00 * 2.0
+
+    def test_cache_ttl_breakdown_real_session_shape(self):
+        """The exact shape observed in a live CC transcript: all 1h, 5m zero."""
+        usage = {
+            "input_tokens": 18,
+            "output_tokens": 160,
+            "cache_read_input_tokens": 31_312,
+            "cache_creation_input_tokens": 31_463,
+            "cache_creation": {
+                "ephemeral_5m_input_tokens": 0,
+                "ephemeral_1h_input_tokens": 31_463,
+            },
+        }
+        cost = cost_tracker.compute_cost(usage, "claude-haiku-4-5")
+        expected = (
+            18 * 1.00 + 160 * 5.00 + 31_312 * 1.00 * 0.10 + 31_463 * 1.00 * 2.0
+        ) / 1e6
+        assert abs(cost - expected) < 1e-12
+
+    def test_cache_breakdown_non_dict_falls_back_to_flat(self):
+        usage = {"cache_creation_input_tokens": 1_000_000, "cache_creation": "junk"}
+        cost = cost_tracker.compute_cost(usage, "claude-opus-4-8")
+        assert cost == 5.00 * 1.25
+
     def test_empty_usage_is_zero(self):
         assert cost_tracker.compute_cost({}, "claude-opus-4-8") == 0.0