@ai-sdk/openai 2.0.136 → 2.0.137

package/dist/internal/index.js

@@ -2000,7 +2000,12 @@ var openaiTranscriptionResponseSchema = (0, import_provider_utils15.lazyValidato
           compression_ratio: import_v410.z.number(),
           no_speech_prob: import_v410.z.number()
         })
-      ).nullish()
+      ).nullish(),
+      // Lenient: an unexpected usage shape must not fail the transcription.
+      usage: import_v410.z.object({
+        type: import_v410.z.string().nullish(),
+        seconds: import_v410.z.number().nullish()
+      }).nullish().catch(void 0)
     })
   )
 );
@@ -2133,10 +2138,10 @@ var OpenAITranscriptionModel = class {
         prompt: openAIOptions.prompt,
         // https://platform.openai.com/docs/api-reference/audio/createTranscription#audio_createtranscription-response_format
         // prefer verbose_json to get segments for models that support it
-        response_format: [
-          "gpt-4o-transcribe",
-          "gpt-4o-mini-transcribe"
-        ].includes(this.modelId) ? "json" : "verbose_json",
+        // gpt-4o(-mini)-transcribe (incl. snapshots) and gpt-transcribe reject verbose_json:
+        response_format: ["gpt-4o-transcribe", "gpt-4o-mini-transcribe"].includes(
+          this.modelId
+        ) || this.modelId.startsWith("gpt-4o-mini-transcribe-") || this.modelId === "gpt-transcribe" || this.modelId.startsWith("gpt-transcribe-") ? "json" : "verbose_json",
         temperature: openAIOptions.temperature,
         timestamp_granularities: openAIOptions.timestampGranularities
       };
@@ -2158,7 +2163,7 @@ var OpenAITranscriptionModel = class {
     };
   }
   async doGenerate(options) {
-    var _a2, _b, _c, _d, _e, _f, _g, _h;
+    var _a2, _b, _c, _d, _e, _f, _g, _h, _i, _j;
     const currentDate = (_c = (_b = (_a2 = this.config._internal) == null ? void 0 : _a2.currentDate) == null ? void 0 : _b.call(_a2)) != null ? _c : /* @__PURE__ */ new Date();
     const { formData, warnings } = await this.getArgs(options);
     const {
@@ -2192,7 +2197,8 @@ var OpenAITranscriptionModel = class {
         endSecond: word.end
       }))) != null ? _g : [],
       language,
-      durationInSeconds: (_h = response.duration) != null ? _h : void 0,
+      // Duration-billed json responses report the duration only in `usage`:
+      durationInSeconds: (_j = response.duration) != null ? _j : ((_h = response.usage) == null ? void 0 : _h.type) === "duration" ? (_i = response.usage.seconds) != null ? _i : void 0 : void 0,
       warnings,
       response: {
         timestamp: currentDate,