vercel/ai · error · UnsupportedFunctionalityError

non-streaming transcription with ${this.modelId}

Error message

non-streaming transcription with ${this.modelId}

What it means

Realtime transcription models (detected via isRealtimeTranscriptionModelId, e.g. gpt-4o-transcribe realtime variants) only support streaming output. Calling doGenerate on OpenAITranscriptionModel with such a model id throws UnsupportedFunctionalityError because there is no non-streaming REST endpoint for it.

Source

Thrown at packages/openai/src/transcription/openai-transcription-model.ts:241

            }
          } else {
            formData.append(key, String(value));
          }
        }
      }
    }

    return {
      formData,
      warnings,
    };
  }

  async doGenerate(
    options: OpenAITranscriptionCallOptions,
  ): Promise<Awaited<ReturnType<TranscriptionModelV4['doGenerate']>>> {
    if (isRealtimeTranscriptionModelId(this.modelId)) {
      throw new UnsupportedFunctionalityError({
        functionality: `non-streaming transcription with ${this.modelId}`,
      });
    }

    const currentDate = this.config._internal?.currentDate?.() ?? new Date();
    const { formData, warnings } = await this.getArgs(options);

    const {
      value: response,
      responseHeaders,
      rawValue: rawResponse,
    } = await postFormDataToApi({
      url: this.config.url({
        path: '/audio/transcriptions',
        modelId: this.modelId,
      }),
      headers: combineHeaders(this.config.headers?.(), options.headers),
      formData,

View on GitHub (pinned to 69428b1f8b)

Solutions

Use the streaming API (streamTranscription/doStream) with realtime transcription models
Switch the model id to a REST-supported transcription model (e.g. whisper-1, gpt-4o-transcribe) for non-streaming use
Branch your code on the model id to select the right call style

Example fix

// before
await transcribe({ model: openai.transcription('gpt-4o-mini-transcribe-realtime'), audio })
// after
const result = streamTranscribe({ model: openai.transcription('gpt-4o-mini-transcribe-realtime'), audio })

Defensive patterns

Strategy: type-guard

Validate before calling

import { isRealtimeTranscriptionModelId } from '@ai-sdk/openai'; // or local copy
if (isRealtimeTranscriptionModelId(modelId)) {
  // use streaming API instead of transcribe()
}

Type guard

function supportsNonStreamingTranscription(modelId: string): boolean {
  return !isRealtimeTranscriptionModelId(modelId);
}

Try / catch

try {
  result = await transcribe({ model, audio });
} catch (e) {
  if (UnsupportedFunctionalityError.isInstance(e) && e.message.includes('non-streaming transcription')) {
    // switch to the streaming transcription API for this model
  } else throw e;
}

Prevention

Keep a mapping of which transcription models support REST vs realtime
Route realtime model ids to streamTranscribe in your abstraction layer
Read the model's provider docs before switching model ids in shared config

When it happens

Trigger: Calling generateText-style transcription (transcribe(...)/doGenerate) with a realtime transcription model id; switching a model id string from a REST transcription model to a realtime one without changing the call to its streaming API.

Common situations: Swapping model ids in config (e.g. 'whisper-1' -> realtime model) while keeping the non-streaming transcribe() call; templated pipelines that always use doGenerate; misunderstanding that realtime models can also do one-shot transcription.

Related errors

AI-assisted analysis of vercel/ai@69428b1f8b (2026-08-30). Data as JSON: /api/errors/12333528768028e0. Report an issue: GitHub.