AI News, 21 October: Eleven Realtime, a retail agent demo and facial performance control

01Eleven Realtime combines the moving parts of a voice agent

ElevenLabs launched Eleven Realtime on 20 October. Its two speech-to-speech models combine listening, turn-taking, language reasoning and spoken replies in one voice-agent choice. Realtime v1 Mini targets everyday conversations, such as an interactive tutor or character. V1 Max is designed for longer instructions and multi-step tool use, such as a support agent that must check an order before answering. Both support more than 90 languages and voices from ElevenLabs' library. The practical gain is less work selecting and tuning separate speech recognition, language and voice models; it does not remove the need to design the agent's instructions or test its behavior.

In ElevenAgents, a builder can select Realtime voice mode on an agent and keep its prompt, tools, knowledge base and channels. A shop could start with Mini for product questions, then compare Max on a support flow that needs several tool calls and careful policy handling. Developers can also use the separate Realtime API through a WebSocket client compatible with OpenAI's Realtime event format. That API route is useful for a custom app, but it does not automatically include the ElevenAgents knowledge base or channels. ElevenLabs reports fast first audio in its benchmark presentation; XANTRANS has not measured latency or listened to these models independently.

Access is rolling out by workspace. If the Voice mode control is missing, that workspace cannot select Realtime yet. ElevenLabs recommends retaining Custom mode when an agent needs a particular language or speech model, Bedrock, managed-setting overrides or transfers to Custom-mode agents. Mini is unavailable for Zero Retention Mode and HIPAA use; Max supports those requirements in US and EU environments but is unavailable in India and Singapore data-residency environments. Realtime uses the normal ElevenAgents call-minute rate plus model token charges, with Max costing more per token than Mini. Check the shown charges for the intended workflow before treating this as a cheaper agent.

02OpenAI's retailer demo links inspiration, a visual preview and order changes

On 21 October, OpenAI published a shopping-agent demonstration set inside Vesper & Alder, a fictional home-goods retailer. A shopper asks by voice for a warmer reading nook and shares a room photo. The agent suggests furniture, generates a view of the proposed arrangement, revises the selection to stay under a $250 budget, and later helps change an order from store pickup to delivery. That sequence shows the appeal of one conversation covering inspiration, product choice and service after purchase. It is a designed example, not a live retailer case study or a newly announced consumer Shopping Agent product.

A real retailer could use the pattern to connect a visual planning assistant to its catalog and order system, letting a customer check whether the suggested room also fits a price ceiling. The demo does not establish that this complete flow is available as a turnkey OpenAI feature, or that every merchant can offer checkout in ChatGPT. OpenAI already introduced ChatGPT shopping research in November 2025 and richer product discovery in March 2026. Its current commerce documents require approved-partner access for product-feed onboarding and a separate checkout integration. Shoppers should confirm final prices and availability with the seller.

03EmotionControl turns a filmed expression into a reusable motion guide

Filmmaker Sirio Berati published EmotionControl's open-source workflow on 14 October. It takes footage of one visible face and creates a clay-like mask video carrying facial motion and source audio. A creator can use that mask with a separate appearance reference in a video generator, giving it a timed performance to follow. For a character close-up, this could guide a smile or hesitation more directly than text alone. The repository produces the mask and comparison files; it does not include a generator or deliver a finished shot.

This is a workflow discovery, not a new 21 October release or a cleared client pipeline. Setup requires Python, FFmpeg and separately obtained SPECTRE and FLAME assets. Although the wrapper code is MIT licensed, those dependencies have noncommercial restrictions, so the full stack cannot be assumed safe for commercial use. The repository tracks one face per frame; turns, blinks and occlusions can disrupt the mask. Source detail may retain likeness. Use footage with performer rights and inspect both the mask and final generation. XANTRANS has not run it.