🍭 Mistral puts AI phone agents on a 4¢ meter ☎️

Your next AI test belongs in the calendar.

Good morning. Mistral’s new voice agents can book appointments, while Google wants Gemini to remember your business’s visual style. Both move AI closer to everyday operations. Before connecting either, pick one result you can check: the right reservation, the right logo, the right receipt total. That is a more useful trial than another impressive conversation.

Today’s Menu 🥠

• Mistral: voice agents start at four cents a minute

• Google: Gemini wants to remember your brand

• Fast Snacks: OpenAI’s biotech acquisition, AI designed phages, Gemini in cars and locked workplace chats

• Toolbox: cache economics, model routing and PyTorch on TPUs

• Image quiz: which floating robot is real?

Wednesday, October 21 · Read Time: 5 minutes

The Big Bite · Mistral
📞 Four Cents a Minute. Can It Finish the Call?

The Bite: Answering the phone is easy. Getting the booking right is the business. Mistral’s Voxtral 3 launch on Tuesday packages speech recognition, a reasoning model, tool calls and speech generation into a voice agent SDK. Mistral says developers can connect it to business systems to check bills or book appointments.

The published voice agent API rate is $0.04 per minute with Mistral Small 4, or $0.08 with another reasoning model. That is an API price, not a complete phone operation’s cost. API access is available now; private cloud and on premises deployments require contacting Mistral.

Why it Bites: The useful shift is from a voice that answers questions to one that can act on them. Mistral reports speech generation in more than 35 languages, but a pleasant voice alone won’t fix a reservation saved for the wrong Tuesday.

Our take: Before routing customers through it, run five messy practice calls: a changed date, a misspelled name, an interruption. Then inspect the actual calendar. Measure correct bookings and necessary human handoffs. A fluent transcript is only half the receipt.

Voices.com
🌐 Live webinar: how to train voice AI that sounds natural

Full-duplex voice models listen and speak at the same time, but only if they’re trained on data that sounds like real conversation.

Open-source and synthetic datasets rarely do. Join Voices and AlphaSignal on Tuesday, October 27 at 1 PM ET as Sully Goolamnobee, Product Manager, AI Voice, covers the three requirements for full-duplex training data and how custom conversational datasets are built to meet them.

Live Q&A included.

Second Bite · Google
🎨 Gemini Wants to Remember Your Brand

The Bite: Google announced a Gemini hub for small businesses that brings marketing, financial analysis, and connected apps into one place. Upload a website, brand kit, fonts, or colors, and Google says Gemini will retain that visual style for future images made with Nano Banana 2.1. The rollout begins globally in the coming days for all Gemini app users.

Why it Bites: The useful part is less repeated setup. A shop owner could prepare a launch graphic, then make the next promotion without explaining the brand again. Connections include Shopify, Canva, Xero, and Google Workspace. Gemini can also turn receipts and spreadsheets into financial graphics and interactive models.

Our take: Test the memory before handing it your marketing calendar. Generate three assets for the same product and check whether the colors, fonts, and logo stay consistent. For financial work, compare its totals with the original records. Saving setup time matters. Quietly introducing a brand mistake or accounting error does, too.

Fast Snacks 🥡

🧬 OpenAI Buys the Work Between Discoveries
OpenAI acquired Orchestra Bio, a platform for coordinating drug development, and is combining its team and technology with Rosalind. Think tracking studies, spotting delays and helping people decide what happens next. A promising molecule still needs a complicated journey to become medicine. This acquisition targets that journey, although OpenAI has not demonstrated how much faster it will become.

🦠 AI Designs Reach a Patient
Radical Numerics says phages designed with its Omnii model were administered to one patient with an infection resistant to antibiotics. Phages are viruses that infect bacteria. The treatment used a physician supervised compassionate use pathway outside a clinical trial. The patient’s outcomes remain private, so this is a reported step into human treatment, with effectiveness still an open question.

🚘 Gemini Brings Your Context Into the Car
Gemini’s optional Personal Intelligence started rolling out on Android Auto Tuesday, letting it retrieve details such as flight information from connected Google apps. Cars with Google built in follow in coming months. Access requires an adult personal account and excludes the EEA, Switzerland, UK and Nigeria. Useful for finding your booking, provided you choose to connect that information.

🔐 A Locked Chat Still Leaves a Record
ChatGPT Enterprise and Edu are gradually adding locked conversations. A PIN or supported device authentication opens them; they disappear from ordinary history and search without being deleted. That can keep a sensitive conversation out of casual view. Your organization’s retention rules still apply, and locked chats remain in compliance logs. A useful privacy layer, with an important workplace boundary.

Toolbox
🧰 Put the next step to work

OpenAI Prompt Caching: For developers repeatedly sending the same context, the Responses API now offers a 12 hour minimum cache lifetime. Writes cost 2.5 times ordinary input token pricing; that is the total write rate, not a surcharge. Savings depend on reuse. Replay a typical day of requests and compare the total bill before switching.

Respan Gateway: For developers juggling AI providers, this Rust gateway puts routing, fallbacks and data redaction into plugins. It is open source and prealpha, so expect changes. You still supply provider keys and pay model usage costs. Start locally with two providers and deliberately trigger a failed request to see whether the fallback preserves your required features.

TorchTPU: For developers with Google TPU access, this Apache 2.0 PyTorch backend supports model training and compilation on those chips. Try a provided Llama or ResNet example on a TPU virtual machine before moving a workload. Compute costs are separate. We found it this week, but its release date is unverified; this is a useful discovery.

Test Your Skill
Snack Quiz: Choose The Real Image 🔍

A floating robot seems to have borrowed somebody’s legs. One image is an archival photograph. The other is an independent AI recreation.

Original photograph: NASA, ISS036 E 013254, June 28, 2013. Source and answer reveal. No NASA endorsement.