fix(pricing): add cache_read_input_token_cost to 8 deepseek models

8 deepseek and openrouter/deepseek models declared only the legacy
`input_cost_per_token_cache_hit` field, but the cost calculator
(`litellm.litellm_core_utils.llm_cost_calc.utils._get_token_base_cost`)
only reads `cache_read_input_token_cost`. As a result, prompt-cache-hit
tokens were billed at $0 instead of the cache-read rate for:

- deepseek/deepseek-coder
- deepseek/deepseek-r1
- deepseek/deepseek-v3.2
- openrouter/deepseek/deepseek-chat-v3.1
- openrouter/deepseek/deepseek-v3.2
- openrouter/deepseek/deepseek-v3.2-exp
- openrouter/deepseek/deepseek-r1
- openrouter/deepseek/deepseek-r1-0528

This patch adds `cache_read_input_token_cost` with the same value as
the existing legacy field to both pricing JSONs, matching the
convention already used by `deepseek/deepseek-chat`,
`deepseek/deepseek-reasoner`, `deepseek/deepseek-v3`, and others.

Adds a regression test that parametrizes over the 8 models and an
end-to-end check that a cache hit on `deepseek/deepseek-r1` is billed
at the cache-read rate via `generic_cost_per_token`.

litellm/model_prices_and_context_window_backup.json

@@ -15333,6 +15333,7 @@
         "supports_tool_choice": true
     },
     "deepseek/deepseek-coder": {
+        "cache_read_input_token_cost": 1.4e-08,
         "input_cost_per_token": 1.4e-07,
         "input_cost_per_token_cache_hit": 1.4e-08,
         "litellm_provider": "deepseek",
@@ -15347,6 +15348,7 @@
         "supports_tool_choice": true
     },
     "deepseek/deepseek-r1": {
+        "cache_read_input_token_cost": 1.4e-07,
         "input_cost_per_token": 5.5e-07,
         "input_cost_per_token_cache_hit": 1.4e-07,
         "litellm_provider": "deepseek",
@@ -15402,6 +15404,7 @@
         "supports_tool_choice": true
     },
     "deepseek/deepseek-v3.2": {
+        "cache_read_input_token_cost": 2.8e-08,
         "input_cost_per_token": 2.8e-07,
         "input_cost_per_token_cache_hit": 2.8e-08,
         "litellm_provider": "deepseek",
@@ -30615,6 +30618,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-chat-v3.1": {
+        "cache_read_input_token_cost": 2e-08,
         "input_cost_per_token": 2e-07,
         "input_cost_per_token_cache_hit": 2e-08,
         "litellm_provider": "openrouter",
@@ -30630,6 +30634,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-v3.2": {
+        "cache_read_input_token_cost": 2.8e-08,
         "input_cost_per_token": 2.8e-07,
         "input_cost_per_token_cache_hit": 2.8e-08,
         "litellm_provider": "openrouter",
@@ -30645,6 +30650,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-v3.2-exp": {
+        "cache_read_input_token_cost": 2e-08,
         "input_cost_per_token": 2e-07,
         "input_cost_per_token_cache_hit": 2e-08,
         "litellm_provider": "openrouter",
@@ -30660,6 +30666,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-r1": {
+        "cache_read_input_token_cost": 1.4e-07,
         "input_cost_per_token": 5.5e-07,
         "input_cost_per_token_cache_hit": 1.4e-07,
         "litellm_provider": "openrouter",
@@ -30675,6 +30682,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-r1-0528": {
+        "cache_read_input_token_cost": 1.4e-07,
         "input_cost_per_token": 5e-07,
         "input_cost_per_token_cache_hit": 1.4e-07,
         "litellm_provider": "openrouter",

model_prices_and_context_window.json

@@ -15333,6 +15333,7 @@
         "supports_tool_choice": true
     },
     "deepseek/deepseek-coder": {
+        "cache_read_input_token_cost": 1.4e-08,
         "input_cost_per_token": 1.4e-07,
         "input_cost_per_token_cache_hit": 1.4e-08,
         "litellm_provider": "deepseek",
@@ -15347,6 +15348,7 @@
         "supports_tool_choice": true
     },
     "deepseek/deepseek-r1": {
+        "cache_read_input_token_cost": 1.4e-07,
         "input_cost_per_token": 5.5e-07,
         "input_cost_per_token_cache_hit": 1.4e-07,
         "litellm_provider": "deepseek",
@@ -15402,6 +15404,7 @@
         "supports_tool_choice": true
     },
     "deepseek/deepseek-v3.2": {
+        "cache_read_input_token_cost": 2.8e-08,
         "input_cost_per_token": 2.8e-07,
         "input_cost_per_token_cache_hit": 2.8e-08,
         "litellm_provider": "deepseek",
@@ -30690,6 +30693,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-chat-v3.1": {
+        "cache_read_input_token_cost": 2e-08,
         "input_cost_per_token": 2e-07,
         "input_cost_per_token_cache_hit": 2e-08,
         "litellm_provider": "openrouter",
@@ -30705,6 +30709,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-v3.2": {
+        "cache_read_input_token_cost": 2.8e-08,
         "input_cost_per_token": 2.8e-07,
         "input_cost_per_token_cache_hit": 2.8e-08,
         "litellm_provider": "openrouter",
@@ -30720,6 +30725,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-v3.2-exp": {
+        "cache_read_input_token_cost": 2e-08,
         "input_cost_per_token": 2e-07,
         "input_cost_per_token_cache_hit": 2e-08,
         "litellm_provider": "openrouter",
@@ -30735,6 +30741,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-r1": {
+        "cache_read_input_token_cost": 1.4e-07,
         "input_cost_per_token": 5.5e-07,
         "input_cost_per_token_cache_hit": 1.4e-07,
         "litellm_provider": "openrouter",
@@ -30750,6 +30757,7 @@
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-r1-0528": {
+        "cache_read_input_token_cost": 1.4e-07,
         "input_cost_per_token": 5e-07,
         "input_cost_per_token_cache_hit": 1.4e-07,
         "litellm_provider": "openrouter",

tests/test_litellm/test_deepseek_cache_read_pricing.py

@@ -0,0 +1,110 @@
+"""
+Regression test for DeepSeek and OpenRouter/DeepSeek models that historically
+only declared ``input_cost_per_token_cache_hit`` in the pricing JSON.
+
+The cost calculator (``litellm.litellm_core_utils.llm_cost_calc.utils``) only
+reads ``cache_read_input_token_cost`` for prompt-cache-hit pricing.  When a
+model defines ``input_cost_per_token_cache_hit`` without the canonical key,
+cache-hit tokens are billed at $0 because the cache-read cost resolves to
+``None`` -> ``0.0``.
+
+This test ensures the two keys are kept in sync for the affected models and
+that a representative model actually bills cache-hit tokens at the cache rate.
+"""
+
+import os
+import sys
+
+import pytest
+
+sys.path.insert(0, os.path.abspath("../.."))
+
+import litellm
+from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_token
+from litellm.types.utils import PromptTokensDetailsWrapper, Usage
+
+# Models that previously only carried ``input_cost_per_token_cache_hit``.
+# Each must also carry ``cache_read_input_token_cost`` with an identical value.
+DEEPSEEK_CACHE_HIT_MODELS = [
+    "deepseek/deepseek-coder",
+    "deepseek/deepseek-r1",
+    "deepseek/deepseek-v3.2",
+    "openrouter/deepseek/deepseek-chat-v3.1",
+    "openrouter/deepseek/deepseek-v3.2",
+    "openrouter/deepseek/deepseek-v3.2-exp",
+    "openrouter/deepseek/deepseek-r1",
+    "openrouter/deepseek/deepseek-r1-0528",
+]
+
+
+@pytest.fixture(autouse=True)
+def _use_local_model_cost_map(monkeypatch):
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    litellm.get_model_info.cache_clear()
+    yield
+    litellm.get_model_info.cache_clear()
+
+
+@pytest.mark.parametrize("model", DEEPSEEK_CACHE_HIT_MODELS)
+def test_cache_read_input_token_cost_present(model):
+    """``cache_read_input_token_cost`` must be present and equal to the legacy
+    ``input_cost_per_token_cache_hit`` field."""
+    info = litellm.model_cost[model]
+    legacy = info.get("input_cost_per_token_cache_hit")
+    canonical = info.get("cache_read_input_token_cost")
+    assert legacy is not None, f"{model} is missing input_cost_per_token_cache_hit"
+    assert canonical is not None, (
+        f"{model} is missing cache_read_input_token_cost; cache-hit tokens would "
+        f"be billed at $0 because the cost calculator only reads the canonical key."
+    )
+    assert canonical == legacy, (
+        f"{model} cache_read_input_token_cost ({canonical}) must match input_cost_per_token_cache_hit ({legacy})"
+    )
+
+
+def test_deepseek_r1_cache_hit_billed_at_cache_rate():
+    """End-to-end check that a cache hit on deepseek/deepseek-r1 is billed at
+    the cache-read rate instead of the regular input rate."""
+    model = "deepseek/deepseek-r1"
+    info = litellm.model_cost[model]
+
+    prompt_tokens = 1000
+    cached_tokens = 800
+    text_tokens = prompt_tokens - cached_tokens
+    completion_tokens = 100
+
+    usage = Usage(
+        prompt_tokens=prompt_tokens,
+        completion_tokens=completion_tokens,
+        total_tokens=prompt_tokens + completion_tokens,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            cached_tokens=cached_tokens,
+            text_tokens=text_tokens,
+        ),
+    )
+
+    input_cost, output_cost = generic_cost_per_token(
+        model=model,
+        usage=usage,
+        custom_llm_provider="deepseek",
+    )
+
+    expected_input_cost = (
+        info["input_cost_per_token"] * text_tokens + info["cache_read_input_token_cost"] * cached_tokens
+    )
+    expected_output_cost = info["output_cost_per_token"] * completion_tokens
+
+    assert abs(input_cost - expected_input_cost) < 1e-12, (
+        f"input cost mismatch: got {input_cost}, expected {expected_input_cost}"
+    )
+    assert abs(output_cost - expected_output_cost) < 1e-12, (
+        f"output cost mismatch: got {output_cost}, expected {expected_output_cost}"
+    )
+
+    # Sanity check: regression scenario (cache_read_input_token_cost missing)
+    # would have billed cached tokens at the full input rate.
+    naive_full_input_cost = info["input_cost_per_token"] * prompt_tokens
+    assert input_cost < naive_full_input_cost, (
+        "cache-hit tokens were not discounted; cache_read_input_token_cost is likely missing from this model entry."
+    )