😺 OpenAI Makes Voice Apps Human (Finally)

Welcome, humans.

Heads up: Anthropic just pulled a policy switcheroo that every Claude user needs to know about. The company updated their terms to flip from “opt-in” to “opt-out” for using your conversations to train future AI models.

Here's the deal: starting September 28, 2025, Anthropic can use your chats with Claude to improve their AI unless you manually opt out in your privacy settings. Before this change, they needed your permission first. Now? They're assuming you're cool with it unless you say otherwise (they also changed their terms of use, so check that too).

The good news is opting out won't break Claude—you'll still get all the same features and functionality. But if you prefer keeping your conversations private, (especially those “how many tablespoons are in a cup again?” queries you ask EVERY DAY), you'll want to head to Claude's settings and toggle off “Help improve Claude.”

Here’s what happened in AI today:

• OpenAI launched GPT Real-time API for human-sounding voice apps.
• Microsoft released MAI-Voice-1 and began testing MAI-1-preview publicly.
• Meta announced AI NPCs coming to Horizon Worlds.
• Agility Robotics trained AI to control their Digit humanoid robot.

Advertise in The Neuron here

OpenAI Just Made It Super Easy to Build Voice Apps That Actually Work

Introducing gpt-realtime in the API

DEEP DIVE: Read the full breakdown on this news here.

OpenAI just made their Realtime API generally available with a brand new speech model called GPT Real-time. Here's why you should care: you can now build voice apps that sound completely human without being a coding wizard.

If you've ever used a voice assistant and thought “this sounds robotic and dumb,” this is OpenAI’s attempt to fix that. The new model can laugh, sigh, switch languages mid-sentence, and actually follow complex instructions without losing track.

Here's what you can actually do with it:

• Build your first voice app: The API now handles everything—you don't need to stitch together multiple services. Just point it at your data and it starts talking naturally about your business, products, or services.
• Connect to real phones: New SIP support means your voice app can answer actual phone calls. Think customer support that doesn't suck, or a personal assistant people can call directly.
• Show it images: Mid-conversation, you can share a screenshot or photo and it'll describe what it sees while keeping the conversation flowing naturally.
• Add real tools: Connect it to your calendar, CRM, or any other service through MCP servers (Connectors). It'll actually use them correctly instead of hallucinating responses.

It costs about $32 per million words spoken to it and $64 per million words it speaks back—which sounds expensive, until you do the math (call it ~$0.75 per hour).

Want to try it? Here’s the links you’ll need:

1. The Realtime API docs to get started.
2. The Realtime prompting guide and Realtime prompting cookbook.
3. The playground link to try it out yourself.

Our take: Voice interfaces are about to get good enough that you'll soon prefer using them to clicking through websites. We're pretty bullish on voice becoming the new interaction framework. But being able to seamlessly switch between voice and text depending on your situation (quiet office versus hitting the road) is where things need to head. We’ve got plenty more to say on this, so check out our deep dive!

FROM OUR PARTNERS

Anything

Welcome to Anything

Anything lets anyone (even you!) launch beautifully designed mobile and web apps from just a prompt. No coding or extra tools required.

Their frontier AI coding agent builds full apps end to end with database, payments, hosting, and one-click App Store submission all included.

Go from idea to live in hours, not months.

Start building with Anything

Prompt Tip of the Day.

As we shared above, OpenAI released their full playbook for prompting gpt-realtime, and it turns out voice AI needs totally different techniques than text models. Here are the game-changing tactics:

Essential Realtime Prompting Techniques:

• Structure your prompt with clear sections: In this order: Role & Objective, Personality & Tone, Tools, Conversation Flow, Safety & Escalation.
• Bullets beat paragraphs: Short bullet points outperform long paragraphs for instruction following.
• Use ALL CAPS for key rules: Capitalizing important instructions makes them stand out
• Convert conditional logic to plain text: Instead of “IF x > 3 THEN ESCALATE,” write “IF MORE THAN THREE FAILURES, THEN ESCALATE.”
• Add tool call preambles: Have the model say “I'm checking that now” before calling functions.
• Control language drift: Pin responses to target languages to prevent unwanted switching.
• Handle unclear audio: Give explicit instructions for background noise and partial words.
• Use sample phrases: Provide examples that show style without being repetitive.
• Add variety rules: The model follows sample phrases so closely that you need explicit variety instructions, or it'll sound like a broken record; for example: “Do not repeat the same sentence twice.”
• Include pronunciation guides: for example, “Pronounce 'SQL' as 'sequel'“ (so adding the phonetic pronunciation) for brand terms.
• Read numbers character-by-character: For phone numbers, codes, etc.
• Use LLMs to review your prompts: Meta-prompt to identify conflicts and ambiguity in what you write.
• Iterate relentlessly: Be prepared to test many small variations—swapping individual words, trying different phrasings for the same instruction, adjusting sentence structure, etc.

Our favorite insight: That last point. Small word swaps like “inaudible” to “unintelligible” can dramatically improve performance. Like GPT-5, the model is incredibly sensitive to precise wording in ways that most text models aren't.

Check out all of our prompt tips of the day from August here!

Treats to Try.

*Asterisk = from our partners. Advertise in The Neuron here.

1. *Chatbase builds customer support agents that automatically answer your customers' questions, update their order details, and escalate complex issues to human agents when needed.
2. Krea announced it plans to release a real-time video model that turns your paintings, prompts, or webcam into real-time video that updates instantly as you make changes (waitlist only rn).
3. Framer generates websites from your text descriptions using AI, then you customize with drag-and-drop (raised $100M).
4. Keychain matches your product ideas with manufacturers—describe “gluten-free protein bars” and get connected to suppliers who can make it (raised $30M).
5. Parsed trains custom language models specifically for your business workflows that keep learning and improving from every interaction, replacing generic "amnesiac" models with specialists that know your domain inside out and perform 50%+ cheaper and 2-3x faster.
6. EliseAI handles communication with apartment renters—it texts back prospects within seconds to schedule tours, follows up on late rent payments, and submits maintenance (raised $250M).
7. Air organizes your creative files and lets you search them by what's actually in them: you can type “red dress” or “John speaking” to instantly find that specific video clip or photo instead of hunting through folders.
8. Check out the 33 top AI companies that have raised over $100M this year… some of them have even raised two rounds of $100M+ this year.

Around the Horn.

• Microsoft launched MAI-Voice-1 (now powering Copilot Daily/Podcasts and usable in Copilot Labs) and began public testing of MAI-1-preview on LMArena with limited API access (apply here).
• Meta will soon let anyone create lifelike AI NPCs that can hold realistic conversations with players in Horizon Worlds via the Worlds Desktop Editor.
• OpenAI released a series of new features for Codex, including a new code editor extension, an updated command line tool with a better interface, the ability to “hand off” tasks to the cloud and write.
• Samsung TVs are getting Microsoft Copilot, so viewers can say “find me something like The Queen's Gambit but about cooking” (so like, The Bear?? lol) and see instant recommendations with photos and ratings.
• Agility Robotics trained a small “motor cortex” AI in simulation that safely balances and moves its Digit humanoid’s whole body that keeps it balanced and lets it walk, lift heavy objects, and do real-world tasks from prompts, which is a step toward general-purpose robots.

FROM OUR PARTNERS

Build AI Agents with Postman

Design, test, and launch AI agents effortlessly with Postman. Use trusted APIs and no-code tools to build workflows and transform them into agent-ready code. Postman gives you everything you need to design, test and launch AI Agents—all in one platform.

Start Building

Intelligent Insights

1. Totally new to AI? Axios put together this simple AI survival kit that breaks down the core concepts and companies to know.
2. Ethan Mollick writes we are entering the “era of Mass Intelligence”, where access to the smartest AI models is getting widespread and easy enough to use that a billion people (or more) are using them.
3. Martin Alderson broke down the cost of AI inference, or running AI models for people to use, and using some back of the napkin math, argued they are largely profitable (especially the API business) because input tokens cost a lot less than output…and yes, even for software engineers.
4. Astral Star Codex explored what’s going on with AI psychosis, how common it is, and even compiled some survey data to try and figure out whether or not it impacted people with no previous psychosis history or risk factors.
5. Here’s the case for why you should still write and code manually, even in the era of AI.

A Cat’s Commentary.