@ai-sdk/anthropic 3.0.81 → 3.0.82

package/src/convert-anthropic-messages-usage.ts

@@ -5,26 +5,25 @@ import type { JSONObject, LanguageModelV3Usage } from '@ai-sdk/provider';
  *
  * - `compaction` / `message`: executor iterations, billed at executor rates.
  * - `advisor_message`: advisor sub-inference, billed at the advisor model's
- *   rates. The `model` field carries the advisor model ID. Advisor tokens
- *   are NOT rolled into the top-level totals because they bill at a
- *   different rate; inspect this array for advisor cost tracking.
+ *   rates. Advisor tokens are NOT rolled into the top-level totals because
+ *   they bill at a different rate; inspect this array for advisor cost
+ *   tracking.
+ * - `fallback_message`: a server-side fallback attempt that served the turn.
+ *   When present, the top-level usage already reflects the served answer, so
+ *   it is used as-is.
+ *
+ * The `model` field carries the model that produced the iteration. The API
+ * populates it for the per-model attribution cases (the fallback chain and
+ * advisor sub-inferences) and omits it otherwise.
  */
-export type AnthropicUsageIteration =
-  | {
-      type: 'compaction' | 'message';
-      input_tokens: number;
-      output_tokens: number;
-      cache_creation_input_tokens?: number | null;
-      cache_read_input_tokens?: number | null;
-    }
-  | {
-      type: 'advisor_message';
-      model: string;
-      input_tokens: number;
-      output_tokens: number;
-      cache_creation_input_tokens?: number | null;
-      cache_read_input_tokens?: number | null;
-    };
+export type AnthropicUsageIteration = {
+  type: 'compaction' | 'message' | 'advisor_message' | 'fallback_message';
+  model?: string | null;
+  input_tokens: number;
+  output_tokens: number;
+  cache_creation_input_tokens?: number | null;
+  cache_read_input_tokens?: number | null;
+};
 
 export type AnthropicMessagesUsage = {
   input_tokens: number;
@@ -57,10 +56,19 @@ export function convertAnthropicMessagesUsage({
   // and output_tokens exclude compaction usage. Advisor (`advisor_message`)
   // iterations are filtered out: they bill at the advisor model's rates,
   // not the executor's, so they don't belong in the top-level totals.
+  //
+  // A turn served by a server-side fallback is the exception: the served
+  // answer comes from the fallback model, so the executor `message` iteration
+  // is the blocked primary attempt (zero output). The top-level totals already
+  // reflect the fallback answer, so they are used directly.
   let inputTokens: number;
   let outputTokens: number;
 
-  if (usage.iterations && usage.iterations.length > 0) {
+  const servedByFallback = usage.iterations?.some(
+    iter => iter.type === 'fallback_message',
+  );
+
+  if (usage.iterations && usage.iterations.length > 0 && !servedByFallback) {
     const executorIterations = usage.iterations.filter(
       iter => iter.type === 'compaction' || iter.type === 'message',
     );