[OPIK-8242] fix: price models with compact YYYYMMDD date suffixes

DATE_SUFFIX_PATTERN only stripped hyphen-separated dates (-2025-12-17),
but Anthropic ships compact dates on every dated model id
(claude-haiku-4-5-20251001). The date fallback in findModelPrice therefore
never fired for them, so a compact-dated name priced correctly only when it
was present verbatim in the price table. Any name needing normalization fell
through every fallback to DEFAULT_COST and reported $0 -- silently, with
token counts still captured.

Most visible via the alias_of entries: "claude-4-6-opus" -> "claude-opus-4-6"
already exists for reversed family/version ordering, so
anthropic/claude-4.6-opus prices at $5/$25, but
anthropic/claude-4.6-opus-20260205 resolved to $0 because the compact date
could not be stripped before the alias lookup. Also hit new releases before
the daily LiteLLM sync picks them up, and is not Anthropic-specific.

Observed in production: 71,442 of ~131,000 LLM spans in one project priced at
$0, hiding ~$4,470 of spend and inflating a 60-day cost delta to +573,216%
where actual spend was flat (~+3%).

Make both separators optional. Month/day ranges are kept so an arbitrary
8-digit build number is not mistaken for a date and cannot collapse a distinct
model onto another model's price row.

The pricing arithmetic itself was verified correct and is unchanged.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: aswynz

apps/opik-backend/src/main/java/com/comet/opik/domain/cost/CostService.java

@@ -71,7 +71,12 @@ public class CostService {
     public static final String MODEL_PRICES_FILE = "model_prices_and_context_window.json";
     public static final String MODEL_PRICES_OVERRIDES_FILE = "model_prices_overrides.json";
     private static final String BEDROCK_PROVIDER = "bedrock";
-    private static final String DATE_SUFFIX_PATTERN = "-\\d{4}-(0[1-9]|1[0-2])-(0[1-9]|[12]\\d|3[01])$";
+    // Both hyphen-separated (2025-12-17) and compact (20251217) date suffixes. Anthropic ships
+    // compact dates on every dated model id (claude-haiku-4-5-20251001), so without the optional
+    // separators a compact-dated name only ever prices when it is present verbatim in the price
+    // table -- any name needing normalization (dot form, an `alias_of` entry, or a release the
+    // daily LiteLLM sync has not picked up yet) silently fell through to a zero cost.
+    private static final String DATE_SUFFIX_PATTERN = "-\\d{4}-?(0[1-9]|1[0-2])-?(0[1-9]|[12]\\d|3[01])$";
     private static final String VERSION_SUFFIX_PATTERN = ":\\d+$";
     private static final Map<String, BiFunction<ModelPrice, Map<String, Integer>, BigDecimal>> PROVIDERS_CACHE_COST_CALCULATOR = Map
             .ofEntries(
@@ -276,7 +281,11 @@ private static String normalizeModelName(String modelName) {
      * This handles cases where providers return dated model names (e.g., "gpt-5.2-2025-12-17")
      * but the pricing database only has the base model name (e.g., "gpt-5.2").
      *
-     * Date patterns recognized: YYYY-MM-DD (e.g., "2025-12-17") at the end of the model name.
+     * Date patterns recognized at the end of the model name: YYYY-MM-DD (e.g., "2025-12-17")
+     * and the compact YYYYMMDD form (e.g., "20251217") that Anthropic uses for every dated
+     * model id. The separators are optional independently, so partially-separated forms are
+     * tolerated too; month and day are still range-checked, so a plain 8-digit build number
+     * such as "-99999999" is left alone.
      *
      * @param modelName The model name
      * @return Lowercase model name with date suffix removed if present, otherwise lowercase original name

apps/opik-backend/src/test/java/com/comet/opik/domain/cost/CostServiceTest.java

@@ -6,6 +6,7 @@
 import org.junit.jupiter.params.ParameterizedTest;
 import org.junit.jupiter.params.provider.Arguments;
 import org.junit.jupiter.params.provider.MethodSource;
+import org.junit.jupiter.params.provider.ValueSource;
 
 import java.math.BigDecimal;
 import java.util.Map;
@@ -375,6 +376,68 @@ void calculateCost_shouldReturnZeroForUnknownModelWithDateSuffix_issue5018() {
         assertThat(cost).isEqualTo(BigDecimal.ZERO);
     }
 
+    /**
+     * A compact-dated name must price at exactly the rate of the base model it normalizes to,
+     * not merely at some non-zero rate -- that is what distinguishes a real price-table hit from
+     * an accidental one. Every case below was observed in production traffic routed through an
+     * enterprise gateway, which emits Anthropic ids in reversed family/version order with a
+     * compact date. Before the fix each of these resolved to DEFAULT_COST and reported $0.00.
+     */
+    @ParameterizedTest
+    @MethodSource("provideCompactDatedModelNamesWithBaseEquivalent")
+    void calculateCost_compactDateSuffixPricesSameAsBaseModel(String datedModelName, String baseModelName,
+            String provider) {
+        Map<String, Integer> usage = Map.of(
+                "prompt_tokens", 1000,
+                "completion_tokens", 500);
+
+        BigDecimal datedCost = CostService.calculateCost(datedModelName, provider, usage, null);
+        BigDecimal baseCost = CostService.calculateCost(baseModelName, provider, usage, null);
+
+        assertThat(baseCost).isGreaterThan(BigDecimal.ZERO);
+        assertThat(datedCost).isEqualByComparingTo(baseCost);
+    }
+
+    private static Stream<Arguments> provideCompactDatedModelNamesWithBaseEquivalent() {
+        return Stream.of(
+                // Reversed family/version order reaches the price row through an `alias_of` entry,
+                // which is only reachable once the compact date is stripped.
+                Arguments.of("anthropic/claude-4.6-opus-20260205", "claude-opus-4-6", "anthropic"),
+                Arguments.of("anthropic/claude-4.6-sonnet-20260217", "claude-sonnet-4-6", "anthropic"),
+                Arguments.of("anthropic/claude-4.5-haiku-20251001", "claude-haiku-4-5", "anthropic"),
+                // Same path without the provider prefix.
+                Arguments.of("claude-4.6-opus-20260205", "claude-opus-4-6", "anthropic"),
+                // Canonical order, compact date, dot form: reaches the row via dot normalization.
+                Arguments.of("claude-opus-4.6-20260205", "claude-opus-4-6", "anthropic"));
+    }
+
+    @Test
+    void calculateCost_shouldReturnZeroForUnknownModelWithCompactDateSuffix() {
+        Map<String, Integer> usage = Map.of(
+                "prompt_tokens", 1000,
+                "completion_tokens", 500);
+
+        BigDecimal cost = CostService.calculateCost("unknown-model-20251217", "openai", usage, null);
+
+        assertThat(cost).isEqualTo(BigDecimal.ZERO);
+    }
+
+    /**
+     * The month/day ranges in the pattern keep an arbitrary 8-digit build or revision number from
+     * being mistaken for a date and silently collapsing a distinct model onto another model's row.
+     */
+    @ParameterizedTest
+    @ValueSource(strings = {"gpt-5.2-99999999", "gpt-5.2-20251345", "gpt-5.2-12345678"})
+    void calculateCost_shouldNotStripNonDateEightDigitSuffix(String modelName) {
+        Map<String, Integer> usage = Map.of(
+                "prompt_tokens", 1000,
+                "completion_tokens", 500);
+
+        BigDecimal cost = CostService.calculateCost(modelName, "openai", usage, null);
+
+        assertThat(cost).isEqualTo(BigDecimal.ZERO);
+    }
+
     /**
      * Test for issue #5621: LiteLLM OTel model names with provider prefix not found in pricing table.
      *
@@ -965,7 +1028,12 @@ private static Stream<Arguments> provideModelNamesWithDateSuffixes() {
                 Arguments.of("claude-sonnet-4.5", "anthropic"),
                 // 4. Provider prefix + date suffix: prefix stripped first, then date suffix removed
                 Arguments.of("anthropic/claude-sonnet-4.5-2025-12-17", "anthropic"),
-                Arguments.of("openai/gpt-5.2-2025-12-17", "openai"));
+                Arguments.of("openai/gpt-5.2-2025-12-17", "openai"),
+                // 5. Compact YYYYMMDD dates, the form Anthropic actually ships on every dated id.
+                Arguments.of("gpt-5.2-20251217", "openai"),
+                Arguments.of("claude-sonnet-4.5-20251217", "anthropic"),
+                Arguments.of("anthropic/claude-sonnet-4.5-20251217", "anthropic"),
+                Arguments.of("openai/gpt-5.2-20251217", "openai"));
     }
 
     /**