chore(v1.10): price-table + model-alias refresh

calibre_toolkit/models.py

@@ -35,7 +35,7 @@
 # the latest model is a one-line edit when a new generation ships.
 _ALIAS_MAP: dict[str, str] = {
     "fast":   "claude-haiku-4-5-20251001",
-    "latest": "claude-opus-4-7",
+    "latest": "claude-opus-4-8",
     "legacy": "claude-sonnet-4-6",
 }
 
@@ -50,8 +50,12 @@
 # (so anyone who pinned a specific ID for reproducibility keeps working)
 # but logs a one-time warning so the user knows it's time to switch.
 _DEPRECATED_MODELS: dict[str, str] = {
-    # Add entries here as Anthropic retires model IDs. Example shape:
-    #   "claude-sonnet-3-5": "claude-sonnet-4-6",
+    # Per Anthropic's deprecation table as of 2026-06-10.
+    "claude-opus-4-1": "claude-opus-4-8",            # retires 2026-08-05
+    "claude-opus-4-1-20250805": "claude-opus-4-8",   # retires 2026-08-05
+    "claude-opus-4-20250514": "claude-opus-4-8",     # retires 2026-06-15
+    "claude-sonnet-4-20250514": "claude-sonnet-4-6", # retires 2026-06-15
+    "claude-3-haiku-20240307": "claude-haiku-4-5",   # retired 2026-04-19
 }
 
 

calibre_toolkit/usage.py

@@ -117,10 +117,11 @@ def _get(name: str) -> int:
 
 # ── Pricing ──────────────────────────────────────────────────────────────────
 #
-# USD per million tokens. Anthropic's public list prices as of January 2026
-# for the Claude 4.x family. Sources: anthropic.com/pricing. Keyed by model
-# family prefix so a new minor version (claude-sonnet-4-7, 4-8) inherits
-# pricing until a deliberate update.
+# USD per million tokens. Anthropic's public list prices, verified against
+# platform.claude.com/docs/en/about-claude/pricing on 2026-06-10. Keyed by
+# model family prefix, longest match wins — so the per-version Opus 4.5+
+# entries take precedence over the bare "claude-opus-4" fallback, which
+# keeps the old $15/$75 rates for Opus 4.0/4.1-era IDs.
 #
 # When the model is unknown we return None and the summary omits the
 # dollar estimate — token counts are still shown so the user can plug
@@ -134,7 +135,20 @@ class _PriceTable:
     cache_write_per_million: float
 
 
+_OPUS_45_PLUS = _PriceTable(
+    input_per_million=5.00,
+    output_per_million=25.00,
+    cache_read_per_million=0.50,
+    cache_write_per_million=6.25,
+)
+
 _PRICES: dict[str, _PriceTable] = {
+    "claude-opus-4-8": _OPUS_45_PLUS,
+    "claude-opus-4-7": _OPUS_45_PLUS,
+    "claude-opus-4-6": _OPUS_45_PLUS,
+    "claude-opus-4-5": _OPUS_45_PLUS,
+    # Opus 4.0/4.1 (deprecated) — dated IDs like claude-opus-4-20250514
+    # fall through to this prefix.
     "claude-opus-4": _PriceTable(
         input_per_million=15.00,
         output_per_million=75.00,
@@ -148,12 +162,13 @@ class _PriceTable:
         cache_write_per_million=3.75,
     ),
     "claude-haiku-4": _PriceTable(
-        input_per_million=0.80,
-        output_per_million=4.00,
-        cache_read_per_million=0.08,
-        cache_write_per_million=1.00,
+        input_per_million=1.00,
+        output_per_million=5.00,
+        cache_read_per_million=0.10,
+        cache_write_per_million=1.25,
     ),
-    # 3.x legacy support — older but still callable.
+    # 3.x legacy — retired models, kept so replaying old usage logs
+    # still produces a dollar estimate.
     "claude-3-5-sonnet": _PriceTable(
         input_per_million=3.00,
         output_per_million=15.00,

tests/test_usage.py

@@ -138,14 +138,19 @@ def test_resolves_sonnet_4_6_prefix(self):
         assert table.output_per_million == 15.00
 
     def test_resolves_opus_prefix(self):
-        table = _resolve_price_table("claude-opus-4-7")
+        # Opus 4.5+ bills at $5/MTok; the bare claude-opus-4 prefix keeps
+        # the old $15 rate for 4.0/4.1-era IDs.
+        table = _resolve_price_table("claude-opus-4-8")
         assert table is not None
-        assert table.input_per_million == 15.00
+        assert table.input_per_million == 5.00
+        legacy = _resolve_price_table("claude-opus-4-20250514")
+        assert legacy is not None
+        assert legacy.input_per_million == 15.00
 
     def test_resolves_haiku_prefix(self):
         table = _resolve_price_table("claude-haiku-4-5-20251001")
         assert table is not None
-        assert table.input_per_million == 0.80
+        assert table.input_per_million == 1.00
 
     def test_unknown_model_returns_none(self):
         assert _resolve_price_table("gpt-4") is None
@@ -205,9 +210,9 @@ def test_cost_estimate_combines_per_model(self):
         agg = UsageAggregate()
         # 1M Sonnet input → $3.00
         agg.record(TokenUsage(input_tokens=1_000_000), "claude-sonnet-4-6")
-        # 1M Haiku input → $0.80
+        # 1M Haiku input → $1.00
         agg.record(TokenUsage(input_tokens=1_000_000), "claude-haiku-4-5")
-        assert agg.cost_estimate_usd() == pytest.approx(3.80)
+        assert agg.cost_estimate_usd() == pytest.approx(4.00)
 
     def test_cost_estimate_none_when_every_model_unpriced(self):
         agg = UsageAggregate()