test: add regression tests for Claude 3 Haiku CRIS cache pricing

Verifies all 4 Bedrock cross-region entries have correct Anthropic-published
cache prices (write $0.30/M, read $0.03/M) matching the direct entry.

tests/test_litellm/test_claude3_haiku_bedrock_cris_cache_pricing.py

@@ -0,0 +1,100 @@
+"""
+Validate that the four Bedrock cross-region Claude 3 Haiku CRIS entries
+carry the correct Anthropic-published cache pricing ($0.30 write / $0.03 read
+per Mtok), matching the direct `claude-3-haiku-20240307` entry.
+
+Previous values were incorrectly derived from the generic 1.25x write / 0.1x
+read ratios (write $0.3125, read $0.025), which understated cache-read costs
+by 16.7% and overstated cache-write costs by 4.2%.
+
+This test guards against regression of the fix in PR #40368.
+"""
+
+import json
+import os
+
+import pytest
+
+JSON_PATH = os.path.join(
+    os.path.dirname(__file__), "../../model_prices_and_context_window.json"
+)
+
+BACKUP_PATH = os.path.join(
+    os.path.dirname(__file__), "../../litellm/model_prices_and_context_window_backup.json"
+)
+
+# The four CRIS (cross-region inference) Bedrock variants that were fixed
+CRIS_KEYS = [
+    "anthropic.claude-3-haiku-20240307-v1:0",
+    "apac.anthropic.claude-3-haiku-20240307-v1:0",
+    "eu.anthropic.claude-3-haiku-20240307-v1:0",
+    "us.anthropic.claude-3-haiku-20240307-v1:0",
+]
+
+# The direct (non-CRIS) entry that has the correct values — used for comparison
+DIRECT_KEY = "claude-3-haiku-20240307"
+
+EXPECTED_CACHE_CREATION = 3e-07  # $0.30 per M tokens
+EXPECTED_CACHE_READ = 3e-08  # $0.03 per M tokens
+
+
+@pytest.fixture(scope="module")
+def model_data():
+    with open(JSON_PATH) as f:
+        return json.load(f)
+
+
+@pytest.fixture(scope="module")
+def backup_data():
+    with open(BACKUP_PATH) as f:
+        return json.load(f)
+
+
+@pytest.mark.parametrize("key", CRIS_KEYS)
+def test_cris_cache_creation_price(model_data, key):
+    """Each CRIS variant should have the correct cache creation price."""
+    assert key in model_data, f"Missing model entry: {key}"
+    entry = model_data[key]
+    actual = entry.get("cache_creation_input_token_cost")
+    assert actual == EXPECTED_CACHE_CREATION, (
+        f"{key}: expected cache_creation_input_token_cost={EXPECTED_CACHE_CREATION} "
+        f"($0.30/M), got {actual}"
+    )
+
+
+@pytest.mark.parametrize("key", CRIS_KEYS)
+def test_cris_cache_read_price(model_data, key):
+    """Each CRIS variant should have the correct cache read price."""
+    assert key in model_data, f"Missing model entry: {key}"
+    entry = model_data[key]
+    actual = entry.get("cache_read_input_token_cost")
+    assert actual == EXPECTED_CACHE_READ, (
+        f"{key}: expected cache_read_input_token_cost={EXPECTED_CACHE_READ} "
+        f"($0.03/M), got {actual}"
+    )
+
+
+@pytest.mark.parametrize("key", CRIS_KEYS)
+def test_cris_prices_match_direct_entry(model_data, key):
+    """CRIS variants should match the direct Claude 3 Haiku entry."""
+    assert key in model_data, f"Missing CRIS entry: {key}"
+    assert DIRECT_KEY in model_data, f"Missing direct entry: {DIRECT_KEY}"
+    cris = model_data[key]
+    direct = model_data[DIRECT_KEY]
+    assert cris.get("cache_creation_input_token_cost") == direct.get("cache_creation_input_token_cost"), (
+        f"{key} cache_creation mismatch with {DIRECT_KEY}"
+    )
+    assert cris.get("cache_read_input_token_cost") == direct.get("cache_read_input_token_cost"), (
+        f"{key} cache_read mismatch with {DIRECT_KEY}"
+    )
+
+
+@pytest.mark.parametrize("key", CRIS_KEYS)
+def test_backup_matches_main(backup_data, model_data, key):
+    """Backup cost map should match the main cost map for all CRIS entries."""
+    assert key in backup_data, f"Missing CRIS entry in backup: {key}"
+    assert backup_data[key].get("cache_creation_input_token_cost") == EXPECTED_CACHE_CREATION
+    assert backup_data[key].get("cache_read_input_token_cost") == EXPECTED_CACHE_READ
+    assert backup_data[key] == model_data[key], (
+        f"Backup and main differ for {key}"
+    )