fix(catalog): give DeepSeek V4 its own window and price

core/providers/catalog.py

@@ -249,6 +249,28 @@ class ModelInfo:
     "deepseek-r1": ModelInfo(
         "deepseek-r1", 128_000, 65_536, 0.55, 2.19, reasoning=_REASONING_ALWAYS_ON
     ),
+    # DeepSeek V4 — source: https://api-docs.deepseek.com/quick_start/pricing
+    # (read 2026-09-27). Every V4 id used to fall through to the ``deepseek``
+    # family rule below, which got two things wrong:
+    #   * window — V4 is 1M with 384K max output, not V3's 128K / 8K, so
+    #     anything compacting against this catalog fired ~8x too early.
+    #   * price — V3's 0.27/1.10 was inherited by *both* tiers, so a pro/flash
+    #     split was invisible to anything that accounts by model id.
+    # V4 is priced by time of day and a row holds one number, so the **peak**
+    # rate is used: the upper bound is the safe direction for a budget guard.
+    # Off-peak is exactly half.
+    # ``deepseek-flash`` is the vendor's current name for V4.1-Flash;
+    # ``deepseek-v4-flash`` is a retired alias that still resolves and is billed
+    # at the Flash price (vendor footnote 1) — same row, both ids.
+    "deepseek-flash": ModelInfo(
+        "deepseek-flash", 1_000_000, 384_000, 0.30, 1.20, reasoning=_REASONING_TOGGLE
+    ),
+    "deepseek-v4-flash": ModelInfo(
+        "deepseek-v4-flash", 1_000_000, 384_000, 0.30, 1.20, reasoning=_REASONING_TOGGLE
+    ),
+    "deepseek-v4-pro": ModelInfo(
+        "deepseek-v4-pro", 1_000_000, 384_000, 1.32, 3.96, reasoning=_REASONING_TOGGLE
+    ),
     # Alibaba Qwen.
     "qwen3-max": ModelInfo(
         "qwen3-max", 256_000, 32_768, 1.2, 6.0, reasoning=_REASONING_TOGGLE
@@ -325,6 +347,13 @@ class ModelInfo:
     ("kimi-latest", _SEED["kimi-k3"]),
     ("kimi", _SEED["kimi-k2.6"]),
     ("deepseek-r", _SEED["deepseek-r1"]),
+    # V4 tiers are seeded now, so an unseen point release inherits the
+    # **cheapest** tier rather than silently taking v3's rate — over-charging a
+    # cheap model is the failure mode that matters for a budget guard. The
+    # vendor's current name gets its own rule so ``deepseek-flash-*`` cannot
+    # fall to v3 either.
+    ("deepseek-flash", _SEED["deepseek-flash"]),
+    ("deepseek-v4", _SEED["deepseek-v4-flash"]),
     ("deepseek", _SEED["deepseek-v3"]),
     ("qwen", _SEED["qwen3-max"]),
     ("grok", _SEED["grok-4"]),

tests/test_catalog.py

@@ -6,6 +6,8 @@
 import sys
 from pathlib import Path
 
+import pytest
+
 ROOT = Path(__file__).resolve().parents[1]
 if str(ROOT) not in sys.path:
     sys.path.insert(0, str(ROOT))
@@ -90,3 +92,66 @@ def test_snapshot_skips_entries_without_context(tmp_path, monkeypatch):
     snap.write_text(json.dumps({"weird-model": {"cost": {"input": 1.0}}}))
     monkeypatch.setattr(catalog, "_SNAPSHOT", {})
     assert catalog.load_catalog_snapshot(snap) == 0
+
+
+# Vendor page read 2026-09-27: https://api-docs.deepseek.com/quick_start/pricing
+# ``deepseek-flash`` is the current name for V4.1-Flash; ``deepseek-v4-flash`` is
+# a retired alias the vendor still accepts and still bills at the Flash price.
+# Both must carry the vendor row instead of falling through to the
+# ``deepseek`` family rule, which also handed V4 a 128K window.
+@pytest.mark.parametrize("model_id", ["deepseek-flash", "deepseek-v4-flash"])
+def test_deepseek_v4_flash_ids_carry_the_vendor_row(model_id):
+    info = catalog.resolve_model_info(model_id)
+
+    assert info.source == "seed"
+    assert info.context_window == 1_000_000
+    assert info.max_output_tokens == 384_000
+
+
+def test_deepseek_v4_tiers_are_priced_apart():
+    # Regression: both V4 tiers used to fall through to the ``deepseek`` family
+    # rule and take ``deepseek-v3``'s price, so pro and flash were one row in
+    # the cost ledger (found by the layer-④ cost census). V4 is priced by time of
+    # day; the seeded numbers are the peak rate, i.e. the upper bound.
+    flash = catalog.resolve_model_info("deepseek-v4-flash")
+    pro = catalog.resolve_model_info("deepseek-v4-pro")
+    v3 = catalog.resolve_model_info("deepseek-v3")
+
+    assert flash.source == "seed"
+    assert pro.source == "seed"
+    assert pro.input_cost_per_1m > flash.input_cost_per_1m
+    assert pro.output_cost_per_1m > flash.output_cost_per_1m
+    assert flash.input_cost_per_1m != v3.input_cost_per_1m
+
+
+@pytest.mark.parametrize(
+    "spelling",
+    [
+        "deepseek-v4-flash",
+        "deepseek/deepseek-v4-flash",
+        "deepseek-ai/DeepSeek-V4-Flash",
+    ],
+)
+def test_observed_gateway_spellings_fold_onto_one_seed_row(spelling):
+    # All three spellings appear in the gateway log for the same logical model.
+    # Prefix-stripping must fold them onto a single id, or cost accounting
+    # fragments by spelling instead of by model.
+    info = catalog.resolve_model_info(spelling)
+
+    assert info.id == "deepseek-v4-flash"
+    assert info.source == "seed"
+
+
+@pytest.mark.parametrize(
+    ("model_id", "expected_source"),
+    [
+        ("deepseek-v4-turbo", "family:deepseek-v4"),
+        ("deepseek-flash-turbo", "family:deepseek-flash"),
+    ],
+)
+def test_unseeded_point_release_does_not_inherit_v3(model_id, expected_source):
+    info = catalog.resolve_model_info(model_id)
+    v3 = catalog.resolve_model_info("deepseek-v3")
+
+    assert info.source == expected_source
+    assert info.input_cost_per_1m != v3.input_cost_per_1m