Voice AI just stopped making people wait. On Monday, two companies shipped AI that listens and talks back in about half a second, and one of them did it in more than 90 languages. If your customers send voice notes instead of typing, this is the one you can't ignore. Behind it: ChatGPT got a lock for your private chats (with a catch), Anthropic's draft IPO papers showed what this whole industry really costs, and a pile of smaller moves you can skim in a minute.
Voice AI now answers in half a second
ElevenLabs launched Eleven Realtime, two speech-to-speech models called Mini and Max. One system hears the customer, works out when they've stopped talking, thinks, and speaks back. The company says replies start in roughly half a second; independent measurements it cited put Max at about 0.57 seconds. The models cover more than 90 languages, and the bigger one is pitched straight at customer support.
The same day, Mistral released Voxtral 3: transcription models, a text-to-speech model and a kit for building voice agents that can call tools and plug into phone lines. Mistral says transcription starts at a fraction of a cent per minute and that companies can run it on their own servers.
Why should a shop owner care? Because a lot of your customers don't type. They hold the mic button and talk, especially on WhatsApp and Telegram, and those voice notes sit unanswered longer than text because you have to stop and listen. Two companies racing to make voice fast and cheap means voice answers are about to become normal, not a gimmick. My take: be skeptical of "90 languages" until you hear it handle your customers' accents and slang. Test with real voice notes, not a demo. If you want to see where voice already pays off in a chat inbox, start with what AI hears in your voice messages and what it sells.
One more thing. Faster, more natural AI voices also make fakes easier. We covered how a cloned voice pulled off a huge WhatsApp scam earlier this month. Agree a simple rule with your team: no payment or address change goes through on a voice note alone.
ChatGPT can lock your chats now. It doesn't encrypt them.
OpenAI added Locked Chats to ChatGPT. You can put a conversation behind a four-digit PIN or Face ID / Touch ID, and it moves to its own section, out of your normal history and search. Locked chats even get their own separate memory, so what you say there doesn't leak into your regular chats. It's rolling out on Free, Go, Plus, Pro, Business, Enterprise and Edu.
Now the catch, and OpenAI says it plainly: locking is not end-to-end encryption. It doesn't change how long the chat is kept, doesn't erase what ChatGPT remembered before you locked it, and doesn't change whether your chats can be used to improve the model. Also, if you reset the PIN, every locked chat is deleted for good.
So what does it mean for you? It's a lock on the drawer, not a safe. Good for keeping your supplier prices away from whoever borrows your phone. Not a reason to paste customer phone numbers, addresses or order lists into a chatbot. That rule hasn't changed.
Anthropic's IPO papers show what AI really costs
A draft IPO filing from Anthropic, the company behind Claude, was reviewed by the Wall Street Journal on Monday, and the numbers are wild in both directions. Revenue for the first half of 2026 was $16.3 billion, up from $1.1 billion a year earlier. The net loss for the same six months was $80.7 billion, though reports say about $72 billion of that is a non-cash accounting charge tied to how the company raised money. It also spent close to $9 billion on computing in six months, more than in all of last year, and has signed up for hundreds of billions more over the next decade. Reports say it could list as soon as November.
Why does a seller care about a lab's balance sheet? Because this is the bill behind every AI tool you use. Somebody pays for those data centers, and in the end it's the customer. Expect more of what we've seen this month: free tiers getting smaller, ads showing up inside chatbots, and prices that move. Don't build your whole sales process on a free plan you can't control. Check how a tool charges you before it becomes the thing you can't live without; our breakdown of what an AI assistant really costs and when it pays you back is a good start.
A safety bill that could pause a model's release
Anthropic also put a draft federal AI safety bill in front of US lawmakers. The idea: independent evaluators, registered with the Commerce Department, check new models. If they find a real, imminent risk that more than 50 people could die or more than $1 billion in damage could happen, from a bioweapon, a cyberattack or a model escaping its makers' control, they have 72 hours to report it. Commerce could then ask a judge to halt the release.
It's a proposal, not a law, and critics already point at a weak spot: the labs could pay the evaluators who check them. For sellers it's background, not a to-do. But it tells you where the rules are heading. Governments are starting to ask who checks these systems before they reach your phone, and that question will reach business tools too.
Quick takes
- Gemini rides along in the car. Google is rolling out Personal Intelligence on Android Auto, so drivers can ask Gemini questions that draw on their Gmail and Google Photos, with permission. More of your customers' questions will come from people who can't look at a screen.
- Cheaper repeat work for AI apps. OpenAI now lets developers cache shared context for up to 12 hours and pay the lower cached rate on repeated material. Boring news that tends to show up later as lower prices in the tools built on top.
- DeepSeek open-sources its serving engine. DeepSeek released DeepInfer, the software it says runs its own production service, for free. More competition on the cost of running AI is good for everyone buying it.
- Otter.ai raises $320 million. The meeting-notes app took non-dilutive money that it pays back from marketing results, and says it passed 40 million users. AI that listens and takes notes is now a big, serious business.
- OpenAI buys Orchestra Bio. A drug-development coordination platform, to sit alongside OpenAI's life-sciences tools. Not a seller story, but a sign of how far the big labs are spreading.
What this week means for your shop
The thread through all of it: AI is getting faster at talking and more expensive to build, and the gap between the two lands on whoever uses it. Your customers will expect an answer in seconds, by text or by voice, and they won't care which company made it. Keep the tool you rely on honest about prices and stock, and keep a person close for the moments that need one. That's the job Vardast does: it answers your customers on Instagram, WhatsApp, Telegram and your website in seconds, from your real product information, and hands the conversation to you the moment it needs a human.