What makes an AI assistant feel like it is actually there? Meta's answer, unveiled at Connect, is a face: Muse Realtime Avatar turns its assistant's voice into an animated character you can hold a live conversation with. The rest of the day was quieter and more practical. Perplexity's new Fast Search, built on an in-house engine called Photon, claims to cut agent search costs by 68 per cent, Amazon put its Seller Assistant inside Claude on Bedrock, and Google DeepMind's new head said Gemini 4 is already in post-training. OpenAI, xAI and most of the big labs were silent, so this is an issue about how AI reaches people rather than how clever it gets: through faces, faster plumbing and the tools they already use at work.
Meta Muse, Perplexity Fast Search, Amazon Seller Assistant and Gemini 4 at a glance
| Item | What is new | Status |
|---|---|---|
| Meta Muse Realtime Avatar | Animated, expressive avatar for live conversations with Muse | Starting in the Muse app |
| Perplexity Fast Search | 95 per cent of results in 230ms or less; 68 per cent lower cost per task than its default preset | Available in the Search API |
| Amazon Seller Assistant plugin | Seller memory plus a plugin into third-party AI tools, including Claude | Live in Amazon Quick; beta with Claude on Bedrock |
| Gemini 4 | In post-training refinement, results described as promising | Could ship before year end |
Meta Muse Realtime Avatar gives the Muse assistant a face you can talk to
Meta used its Connect keynote to unveil Muse Realtime Avatar, which turns Muse's real-time voice into an expressive, animated avatar you can hold a live conversation with. If you have tried Muse as a personal agent (it reached the Mac last week) and found it competent but faceless, this closes that gap: conversations now happen with something that looks and reacts as if it is listening, starting inside the Muse app.
It is a meaningful step in the race to make assistants feel present rather than transactional, and it puts pressure on every rival still limited to text or audio-only voice.
Perplexity Fast Search and the Photon engine cut agent search costs by 68 per cent
Perplexity has added Fast Search to its Search API, built on a new in-house retrieval and ranking engine called Photon. The headline figure is that 95 per cent of results come back in 230 milliseconds or less, and Perplexity's own benchmarking shows it undercutting rivals on both speed and cost, with cost per task down 68 per cent against its own default preset across six agent benchmarks.
If you build on Perplexity's API, or you are choosing a search backend for an agent pipeline, this is the kind of infrastructure upgrade that changes your unit economics without touching a single prompt. It follows the company's push to run agents locally on Windows earlier this month. The benchmarks are Perplexity's own, so test against your workload before switching.
Amazon Seller Assistant plugin lets sellers run their storefront from Claude
Amazon has upgraded its Seller Assistant with persistent memory of each seller's business, plus a new plugin that carries that knowledge into third-party AI tools. It launches first with Amazon's own Quick assistant and is already in beta with Anthropic's Claude on Bedrock, so sellers will be able to manage inventory, pricing and account health from inside Claude rather than switching back to Seller Central.
It is a good example of a pattern worth watching: on platforms of Amazon's scale, the contest is no longer only about the model but about who gets first-party plugin access into the workflows people already spend their day in.
Industry themes
The strongest item today is about presence rather than intelligence. Meta's real-time Muse avatars, alongside Google DeepMind's recent Live Avatar work on Gemini, push assistants from something you type or talk to towards something that visibly listens back, which may matter more to daily users than another benchmark point.
Underneath the model headlines, the infrastructure layer is getting a speed war of its own. Perplexity's Fast Search claims on latency and cost per task are aimed squarely at agent builders, for whom search calls are a growing slice of the bill.
And AI keeps arriving inside the tools people already use rather than as a destination of its own, from Amazon's seller plugin for Claude to wider free access to OpenEvidence for clinicians. With Gemini 4 now in post-training, expect the frontier race to pick up again before Christmas.