What's new at Inworld

Realtime TTS-2 Flash voices no longer clip the opening sound of a sentence, so words like “Seven” and “Tomorrow” come through whole.

Decision models return typed answers your code can branch on, through the Decisions API and the Portal playground. The lineup now includes Cloudflare Clef, Clef Flash, Liquid AI D1, OpenAI GPT 6 Luna Decisions, TypeSafe Jev and Jev 1.13.

Generate video through the Video API and the Portal playground with Alibaba Wan 3.0, MiniMax H3 and xAI Grok Imagine Video 1.5 Lite. Supported models can also start from a first frame or reference images.

Run speech to text in our Europe and Asia Pacific regions, so audio stays in the region and people nearby get their transcripts sooner. Regional deployments are on Enterprise plans.

On Builder plans and above, you can pull your usage into your own systems with the Usage API.

Realtime TTS-2 now returns lip-sync timings, so your characters' mouths move in time with their speech.

Muse Glimmer 30B and Muse Spark 1.3 are now available through Inworld Inference. Muse Glimmer 30B is a reasoning model with function calling and structured output, built for multi-step problems where accuracy matters most.

Every API response now includes an x-request-id header, and you can send your own x-inworld-client-request-id to Inference, text to speech and speech to text. Include either one in a support ticket and we can trace the exact request.

With the Inworld CLI 1.3.1 and the Claude Code plugin, your coding agent can synthesize speech, transcribe audio, clone voices, chat with models and search our docs.

Teach text to speech how to say brand names, place names and surnames. Save dictionaries in the Portal under Pronunciations, and set a default for your whole workspace.

Transcripts now label who is speaking, using a new, more accurate model. It works on live audio and on recordings sent through async transcription, as an add-on at $0.02 an hour.

Submit a long recording by upload or URL, and collect the transcript when it is ready.

Create professional-quality custom voices in languages beyond English, including German, Russian and Hindi.

Word timestamps now account for silence, so they line up precisely with the audio.

Speech now keeps its tone and pacing after an SSML pause instead of starting over.

Rate limit errors now say exactly when you can try again, for example “in 5 hours 12 minutes”. Upgrading your plan raises your hourly and daily limits right away.

After you upgrade, your new professional voice cloning limits apply immediately, and a failed request explains what to do next.

Speech to speech sessions now connect reliably, even during heavy traffic.

WAV download is now included on every plan, not just Creator and above.

On Enterprise and Enterprise Trial plans you can now turn on zero data retention in the Portal, so we keep none of your request or response content. On another plan? Talk to us, and we can turn it on for you once your account meets the conditions.