solidify logic

routstr/payment/models.py

@@ -94,7 +94,10 @@ def backfill_cache_pricing(model_id: str, pricing: Pricing) -> Pricing:
     cache reads (DeepSeek hits are 10x cheaper) and undercharges Anthropic
     cache writes (1.25x). litellm ships per-model USD rates keyed by the exact
     OpenRouter id (deepseek/deepseek-chat) or by the bare model name
-    (gpt-4o, claude-sonnet-4-5), so both spellings are tried.
+    (gpt-4o, claude-sonnet-4-5), so both spellings are tried. litellm keys are
+    lowercase, but a generic upstream may report a mixed-case id
+    (``deepseek-ai/DeepSeek-V4-Flash``); an exact match is attempted first, then
+    a case-insensitive fallback so such ids still resolve.
 
     Rates already present (e.g. provided by OpenRouter) are authoritative and
     never overwritten. Unknown models are returned unchanged.
@@ -106,12 +109,26 @@ def backfill_cache_pricing(model_id: str, pricing: Pricing) -> Pricing:
 
     import litellm
 
+    candidates = (model_id, model_id.split("/", 1)[-1])
     info: dict | None = None
-    for key in (model_id, model_id.split("/", 1)[-1]):
+    for key in candidates:
         candidate = litellm.model_cost.get(key)
         if isinstance(candidate, dict):
             info = candidate
             break
+    if info is None:
+        # Case-insensitive fallback: a mixed-case upstream id (e.g.
+        # ``deepseek-ai/DeepSeek-V4-Flash``) won't match litellm's lowercase
+        # keys exactly. Build a lowercased index once and retry.
+        lowered = {c.lower() for c in candidates}
+        for key, candidate in litellm.model_cost.items():
+            if (
+                isinstance(key, str)
+                and key.lower() in lowered
+                and isinstance(candidate, dict)
+            ):
+                info = candidate
+                break
     if info is None:
         return pricing
 
@@ -215,13 +232,29 @@ def _row_to_model(
     )
     top_provider_dict = json.loads(row.top_provider) if row.top_provider else None
 
-    if apply_provider_fee and isinstance(pricing, dict):
-        pricing = {k: float(v) * provider_fee for k, v in pricing.items()}
-
     if isinstance(pricing, dict) and float(pricing.get("request", 0.0)) <= 0.0:
         pricing["request"] = max(pricing.get("request", 0.0), 0.0)
 
     parsed_pricing = Pricing.parse_obj(pricing)
+
+    # Fill missing cache-read/write rates from litellm's cost map BEFORE applying
+    # the provider fee, so they carry the same markup as every other component.
+    # DB-stored override pricing (e.g. generic providers) omits cache rates;
+    # without this, ``_row_to_model`` bills cache reads at the full input rate —
+    # the ``_apply_provider_fee_to_model`` path backfills, but the override path
+    # used for admin-configured providers did not.
+    #
+    # Key on ``forwarded_model_id`` (the actual upstream model name litellm
+    # prices) when set: an alias row (id="local-alias",
+    # forwarded_model_id="deepseek-v4-flash") would otherwise look up the alias
+    # and miss the cache rate.
+    pricing_model_id = getattr(row, "forwarded_model_id", None) or row.id
+    parsed_pricing = backfill_cache_pricing(pricing_model_id, parsed_pricing)
+
+    if apply_provider_fee:
+        parsed_pricing = Pricing.parse_obj(
+            {k: float(v) * provider_fee for k, v in parsed_pricing.dict().items()}
+        )
     model = Model(
         id=row.id,
         name=row.name,

routstr/upstream/deepseek_v4_pricing_shim.py

@@ -3,10 +3,15 @@
 litellm's bundled cost map does not yet ship ``deepseek-v4-flash`` /
 ``deepseek-v4-pro``. Without an entry, ``backfill_cache_pricing`` cannot find a
 ``cache_read_input_token_cost`` and cache reads fall back to the full input
-rate — a ~60% overcharge on cache hits (DeepSeek hits are ~0.2x input).
+rate — a large overcharge on cache hits (DeepSeek V4 hits are ~0.008-0.02x
+input, i.e. cached tokens cost 50-120x less than regular input).
 
 This module injects the missing entries into ``litellm.model_cost`` at startup
-so the existing backfill path resolves them. Rates mirror the open upstream PR
+so the existing backfill path resolves them. Rates mirror the canonical
+``deepseek`` provider entries now in litellm's ``model_prices`` map
+(``input_cost_per_token`` is the cache-*miss* rate;
+``cache_read_input_token_cost`` is the cache-*hit* rate), sourced from
+https://api-docs.deepseek.com/quick_start/pricing via
 https://github.com/BerriAI/litellm/pull/26380 (issue
 https://github.com/BerriAI/litellm/issues/30430).
 
@@ -22,21 +27,23 @@ from ..core import get_logger
 
 logger = get_logger(__name__)
 
-# USD per token. Source: BerriAI/litellm PR #26380.
+# USD per token. Mirrors the canonical ``deepseek`` provider entries in
+# litellm's model_prices map (source: DeepSeek API pricing docs). Keep these in
+# sync with ``litellm.model_cost["deepseek/deepseek-v4-*"]``.
 _DEEPSEEK_V4_RATES: dict[str, dict[str, float]] = {
     "deepseek-v4-flash": {
         "input_cost_per_token": 1.4e-07,
         "output_cost_per_token": 2.8e-07,
-        "cache_read_input_token_cost": 2.8e-08,
+        "cache_read_input_token_cost": 2.8e-09,
         "cache_creation_input_token_cost": 0.0,
-        "input_cost_per_token_cache_hit": 2.8e-08,
+        "input_cost_per_token_cache_hit": 2.8e-09,
     },
     "deepseek-v4-pro": {
-        "input_cost_per_token": 1.74e-06,
-        "output_cost_per_token": 3.48e-06,
-        "cache_read_input_token_cost": 1.4e-07,
+        "input_cost_per_token": 4.35e-07,
+        "output_cost_per_token": 8.7e-07,
+        "cache_read_input_token_cost": 3.625e-09,
         "cache_creation_input_token_cost": 0.0,
-        "input_cost_per_token_cache_hit": 1.4e-07,
+        "input_cost_per_token_cache_hit": 3.625e-09,
     },
 }
 

tests/unit/test_cache_pricing.py

@@ -81,6 +81,21 @@ def test_backfill_strips_vendor_prefix_for_litellm_lookup() -> None:
     assert result.input_cache_read == expected
 
 
+def test_backfill_case_insensitive_lookup() -> None:
+    """A generic upstream may report a mixed-case id
+    (deepseek-ai/DeepSeek-V4-Flash); litellm keys are lowercase. The
+    case-insensitive fallback still resolves the cache rate."""
+    pricing = Pricing(prompt=1.4e-07, completion=2.8e-07)
+
+    result = backfill_cache_pricing("deepseek-ai/DeepSeek-V4-Flash", pricing)
+
+    expected = litellm.model_cost["deepseek-v4-flash"][
+        "cache_read_input_token_cost"
+    ]
+    assert result.input_cache_read == expected
+    assert result.input_cache_read < pricing.prompt  # sanity: it's a discount
+
+
 def test_backfill_fills_cache_write_rate() -> None:
     """Anthropic cache writes cost more than input (1.25x); billing them at
     the input rate undercharges. litellm carries the write rate."""
@@ -136,6 +151,93 @@ def test_provider_fee_applies_to_backfilled_cache_rates() -> None:
     assert adjusted.pricing.prompt == pytest.approx(2.8e-07 * 2.0)
 
 
+def test_row_to_model_backfills_cache_rate() -> None:
+    """The DB-override path (admin-configured providers, e.g. a generic
+    upstream) stores pricing without cache rates. ``_row_to_model`` must
+    backfill them from litellm just like ``_apply_provider_fee_to_model``,
+    otherwise cache reads bill at the full input rate."""
+    import json
+
+    from routstr.core.db import ModelRow
+    from routstr.payment.models import _row_to_model
+
+    row = ModelRow(
+        id="deepseek-v4-flash",
+        name="deepseek-v4-flash",
+        created=0,
+        description="",
+        context_length=1000000,
+        architecture=json.dumps(
+            {
+                "modality": "text",
+                "input_modalities": ["text"],
+                "output_modalities": ["text"],
+                "tokenizer": "unknown",
+                "instruct_type": None,
+            }
+        ),
+        # Stored pricing omits input_cache_read (generic provider never sets it).
+        pricing=json.dumps({"prompt": 1.4e-07, "completion": 2.8e-07}),
+        enabled=True,
+        upstream_provider_id=1,
+    )
+
+    with patch(
+        "routstr.payment.models.sats_usd_price", return_value=5.0e-5
+    ):
+        model = _row_to_model(row, apply_provider_fee=True, provider_fee=1.0)
+
+    litellm_read = litellm.model_cost["deepseek-v4-flash"][
+        "cache_read_input_token_cost"
+    ]
+    assert model.pricing.input_cache_read == pytest.approx(litellm_read)
+    assert model.pricing.input_cache_read < model.pricing.prompt  # a discount
+    assert model.sats_pricing is not None
+    assert model.sats_pricing.input_cache_read > 0
+
+
+def test_row_to_model_backfills_via_forwarded_model_id() -> None:
+    """An alias row (id != forwarded_model_id) must backfill cache rates from
+    the *forwarded* model name — the real upstream model litellm prices —
+    not the alias id, which litellm doesn't know."""
+    import json
+
+    from routstr.core.db import ModelRow
+    from routstr.payment.models import _row_to_model
+
+    row = ModelRow(
+        id="local-alias",  # litellm has no such key
+        name="local-alias",
+        created=0,
+        description="",
+        context_length=1000000,
+        architecture=json.dumps(
+            {
+                "modality": "text",
+                "input_modalities": ["text"],
+                "output_modalities": ["text"],
+                "tokenizer": "unknown",
+                "instruct_type": None,
+            }
+        ),
+        pricing=json.dumps({"prompt": 1.4e-07, "completion": 2.8e-07}),
+        enabled=True,
+        upstream_provider_id=1,
+        forwarded_model_id="deepseek-v4-flash",
+    )
+
+    with patch(
+        "routstr.payment.models.sats_usd_price", return_value=5.0e-5
+    ):
+        model = _row_to_model(row, apply_provider_fee=True, provider_fee=1.0)
+
+    litellm_read = litellm.model_cost["deepseek-v4-flash"][
+        "cache_read_input_token_cost"
+    ]
+    assert model.pricing.input_cache_read == pytest.approx(litellm_read)
+    assert model.pricing.input_cache_read < model.pricing.prompt
+
+
 # ============================================================================
 # calculate_cost — cached tokens billed at cache rates
 # ============================================================================