Claude Fable 5.1: pricing entry with the 0.025x cache-read exception

CHANGELOG.md

@@ -10,6 +10,13 @@ See [docs/RELEASING.md](docs/RELEASING.md) for the release procedure.
 ## [Unreleased]
 
 ### Added
+- Bundled price catalog: `claude-fable-5-1` ($10/$50 per 1M tokens, 1-hour cache
+  write $20, **cache read $0.25 — 0.025x base input, not the 0.1x every other
+  Claude model uses**), verified against the official Anthropic pricing page
+  2026-09-04; the other Claude entries re-verified at the same date. ([#80])
+- Bundled price catalog: `claude-sonnet-5` ($2/$10 per 1M tokens, cache read
+  $0.20 / 1-hour cache write $4), verified against the official Anthropic
+  pricing page 2026-08-24. ([#78])
 - Bundled price catalog: `claude-opus-5` ($5/$25 per 1M tokens, cache read
   $0.50 / 1-hour cache write $10), verified against the official Anthropic
   pricing page 2026-07-24; the other Claude entries re-verified at the same
@@ -644,5 +651,7 @@ Initial release under the **nasde-toolkit** name (rebrand from
 [#73]: https://github.com/NoesisVision/nasde-toolkit/pull/73
 [#74]: https://github.com/NoesisVision/nasde-toolkit/pull/74
 [#77]: https://github.com/NoesisVision/nasde-toolkit/pull/77
+[#78]: https://github.com/NoesisVision/nasde-toolkit/pull/78
+[#80]: https://github.com/NoesisVision/nasde-toolkit/pull/80
 [#81]: https://github.com/NoesisVision/nasde-toolkit/pull/81
 [gh-litellm-2026-04]: https://github.com/BerriAI/litellm/security/advisories/GHSA-xqmj-j6mv-4862

src/nasde_toolkit/pricing.toml

@@ -15,10 +15,14 @@
 # per-step cost accounting on the 2026-07 grid. OpenAI bills no write premium,
 # so gpt entries omit the field (writes fall back to the base input rate).
 #
-# Claude models re-verified 2026-08-24 against the official pricing page (Fable 5,
-# Opus 5, Opus 4.8, Sonnet 4.6 unchanged since 2026-07-13; Sonnet 5 added) —
-# CONFIRM before publishing any cost figures. gpt entries not re-checked since
-# 2026-06-08.
+# Cache reads are 0.1x the base input rate on every Claude model EXCEPT
+# Claude Fable 5.1 (and Mythos 5.1), where they are 0.025x — $0.25/MTok on a
+# $10/MTok input rate. That is a published per-model exception, not a typo.
+#
+# Claude models re-verified 2026-09-04 against the official pricing page (Fable 5,
+# Opus 5, Opus 4.8, Sonnet 5, Sonnet 4.6 unchanged since 2026-08-24; Fable 5.1
+# added) — CONFIRM before publishing any cost figures. gpt entries not re-checked
+# since 2026-06-08.
 
 [models."gpt-5.5"]
 input_per_1m = 5.0
@@ -39,37 +43,45 @@ input_per_1m = 5.0
 output_per_1m = 25.0
 cached_input_per_1m = 0.50
 cache_write_per_1m = 10.0
-as_of = "2026-08-24"
+as_of = "2026-09-04"
 source = "https://platform.claude.com/docs/en/about-claude/pricing"
 
 [models."claude-opus-4-8"]
 input_per_1m = 5.0
 output_per_1m = 25.0
 cached_input_per_1m = 0.50
 cache_write_per_1m = 10.0
-as_of = "2026-08-24"
+as_of = "2026-09-04"
+source = "https://platform.claude.com/docs/en/about-claude/pricing"
+
+[models."claude-fable-5-1"]
+input_per_1m = 10.0
+output_per_1m = 50.0
+cached_input_per_1m = 0.25
+cache_write_per_1m = 20.0
+as_of = "2026-09-04"
 source = "https://platform.claude.com/docs/en/about-claude/pricing"
 
 [models."claude-fable-5"]
 input_per_1m = 10.0
 output_per_1m = 50.0
 cached_input_per_1m = 1.0
 cache_write_per_1m = 20.0
-as_of = "2026-08-24"
+as_of = "2026-09-04"
 source = "https://platform.claude.com/docs/en/about-claude/pricing"
 
 [models."claude-sonnet-5"]
 input_per_1m = 2.0
 output_per_1m = 10.0
 cached_input_per_1m = 0.20
 cache_write_per_1m = 4.0
-as_of = "2026-08-24"
+as_of = "2026-09-04"
 source = "https://platform.claude.com/docs/en/about-claude/pricing"
 
 [models."claude-sonnet-4-6"]
 input_per_1m = 3.0
 output_per_1m = 15.0
 cached_input_per_1m = 0.30
 cache_write_per_1m = 6.0
-as_of = "2026-08-24"
+as_of = "2026-09-04"
 source = "https://platform.claude.com/docs/en/about-claude/pricing"

tests/test_evaluator.py

@@ -282,7 +282,7 @@ def test_assessment_summary_includes_economics(tmp_path: Path) -> None:
     assert summary.token_usage["total_tokens"] == 1_060_000
     # sonnet $3/$15/$0.30 cached (ADR-014): fresh 0.2M*3 + cached 0.8M*0.30 + 0.06M*15 = 1.74
     assert summary.cost_usd == pytest.approx(1.74)
-    assert summary.pricing_as_of == "2026-08-24"
+    assert summary.pricing_as_of == "2026-09-04"
     assert not hasattr(summary, "cost_efficiency")  # removed: arbitrary zero → use Pareto front
     assert not hasattr(summary, "token_efficiency")
 

tests/test_pricing.py

@@ -87,6 +87,31 @@ def test_compute_cost_cache_aware_matches_harbor_accounting() -> None:
     assert cost == pytest.approx(8.841665, abs=0.0005)
 
 
+def test_compute_cost_cache_aware_matches_harbor_accounting_fable_5_1() -> None:
+    # Real trial qoVXn7k (claude-fable-5-1, xhigh): Harbor's own per-step total was
+    # $3.57704175 — it applies the same 0.025x read rate. Pricing reads at 0.1x
+    # like every other Claude model would land at $4.38 (+22% on this token mix).
+    cost = compute_cost_usd(
+        1_160_418,
+        30_522,
+        "claude-fable-5-1",
+        load_pricing(),
+        cache_read_tokens=1_070_967,
+        cache_write_tokens=88_869,
+    )
+    assert cost == pytest.approx(3.57704175, abs=0.0005)
+
+
+def test_fable_5_1_cache_reads_bill_at_the_published_quarter_rate() -> None:
+    # Fable 5.1 reads are 0.025x base input ($0.25 on $10), not the 0.1x every
+    # other Claude model uses — guards the entry against a "consistency" fix.
+    pricing = load_pricing()
+    assert pricing["claude-fable-5-1"].cached_input_per_1m == 0.25
+    # 1M prompt tokens, all cache reads, 0.1M output: 0.25 + 5.0
+    cost = compute_cost_usd(1_000_000, 100_000, "claude-fable-5-1", pricing, cache_read_tokens=1_000_000)
+    assert cost == pytest.approx(5.25)
+
+
 def test_compute_cost_missing_cache_rates_falls_back_to_full_rate(tmp_path: Path) -> None:
     custom = tmp_path / "pricing.toml"
     custom.write_text('[models."bare"]\ninput_per_1m = 2.0\noutput_per_1m = 10.0\n')
@@ -169,6 +194,7 @@ def test_layered_three_layers_compose(tmp_path: Path, empty_user_layer: Path) ->
         "gpt-5.5",
         "gpt-5.4",
         "claude-opus-5",
+        "claude-fable-5-1",
         "claude-fable-5",
         "claude-opus-4-8",
         "claude-sonnet-5",

tests/test_results_exporter.py

@@ -174,7 +174,7 @@ def test_export_includes_token_cost_economics(job_dir: Path, tmp_path: Path) ->
     assert usage["total_tokens"] == 1_060_000
     # sonnet $3/$15/$0.30 cached (ADR-014): fresh 0.2M*3 + cached 0.8M*0.30 + 0.06M*15 = 1.74
     assert metrics["cost_usd"] == pytest.approx(1.74)
-    assert metrics["pricing_as_of"] == "2026-08-24"
+    assert metrics["pricing_as_of"] == "2026-09-04"
     assert "cost_efficiency" not in metrics  # removed: arbitrary zero → use Pareto front
     assert "token_efficiency" not in metrics
     assert metrics["reasoning_effort"] == ""  # fixture set no override