G'day. Five releases from the last day or so, why each one matters, and how to try it. It's a big day for voice: two of today's five are about AI that talks, and two more are free to try. About a five-minute read.

French AI lab Mistral has released Voxtral 3, a family of speech models plus a toolkit for building voice agents that listen, think, call your tools and talk back in real time. The transcription models cover 50+ languages, can tell who said what, and know when someone has finished speaking. The new text-to-speech model speaks 35+ languages and 50+ dialects and changes its emotion and pace to suit the conversation. Mistral says it beats Whisper, ElevenLabs' Scribe v2 and Gemini 3.5 Transcribe on transcription accuracy. You can use it through Mistral's API or run it on your own servers, which matters if your call recordings can't leave the building.

Why it matters: until now, building a phone or web voice agent meant stitching four separate products together. This puts the whole chain in one kit from one supplier, at prices well under the big names.

TRY IT → in the Studio playground on Mistral's site, then through the API. Transcription starts at about A$0.005 a minute (US$0.0035), so an hour of audio costs about 30 cents. Real-time transcription is about A$0.011 a minute (US$0.008), and the voice costs about A$18.57 per million characters (US$13). Mistral: Voxtral 3 · Model details and pricing

Paris start-up Gradium has open-sourced Phonon, a family of tiny text-to-speech models built to run on the device itself, with no internet connection. The English model, Phonon 31m, is just 64.5 MB, and Gradium says it starts speaking in 7 milliseconds on a laptop's ordinary processor. It runs on Windows, Mac and Linux, Android phones, iPhones, a Raspberry Pi 5 and even inside a web page. A slightly bigger 40m version also speaks French, German, Spanish and Portuguese. Your text never leaves the device, and there's no per-word bill.

Why it matters: good-sounding AI voices have meant paying a cloud service for every sentence. This one is free, private and works on a plane.

TRY IT → free. The English Phonon 31m is under the Apache 2.0 licence, so you can use it in commercial apps. The 40m models are free for non-commercial use only. If you have Python, pip install ptts and then ptts --lang en "Hello from Phonon." -o speech.wav will speak your first line, or you can hear it in your browser on Gradium's page. Gradium: Phonon · Launch post · Hugging Face: phonon-31m-en

Start-up Asymptote Labs has officially launched Beacon, along with US$5.5 million in seed funding. Beacon quietly records your sessions in Claude Code, Cursor, Codex, OpenCode and 20+ other coding tools, then picks out what worked, such as the fix for an obscure bug or the right way to run your tests, and turns it into reusable "skills". You review each one, and from then on every agent you use can look it up, whichever tool it lives in. By default your history stays on your own computer.

Why it matters: every time you switch coding tools, or even start a new chat, you end up explaining your project all over again. Beacon is a fix that isn't tied to any one company's agent.

TRY IT → free and open source (MIT licence). The local version and the basic Beacon Cloud dashboard are free. Beacon Cloud Pro, which adds unlimited findings, memory over MCP (the standard way agents connect to outside tools) and real-time alerts, is US$20 a month (about A$29). The quickest start is to paste "Read https://beacon.sh/start.md then set me up on Beacon" into your coding agent. Beacon · GitHub: Asymptote-Labs/agent-beacon

Square Enix's new Final Fantasy launches today on Switch 2, Switch, PS5 and Xbox Series X|S, with the Steam version following on 23 October. It's the series' first game in the "HD-2D" look made famous by Octopath Traveler: pixel-art characters in detailed 3D worlds. It's a tale of crystals with turn-based battles, chocobos, espers and airships, plus cameos from Cloud, Tidus and the Warrior of Light. The story is a rebuilt version of the first arc of the mobile game Final Fantasy Brave Exvius.

Why it matters: it's a proper old-school Final Fantasy for anyone who misses the classics, and the Switch versions mean you can play it on the train.

TRY IT → A$74.95 on Steam, or about A$84.95 for the console versions at Australian stores (some retailers are discounting). Not sure? There's a free demo on Steam first. Square Enix: Final Fantasy Resonance · Steam: Final Fantasy Resonance

New AI lab humans& has released its first model, Ampersand, and it's aimed at a different goal from most. Instead of training a model to finish tasks on its own, humans& trained Ampersand inside simulated groups of people over months of simulated time, scoring it on how much it helped each person towards their long-term goals without stepping on their preferences. Think of three siblings planning their parents' 40th anniversary over five weeks. It's built on the open GLM 5.3 model. humans& is upfront that it can be a bit much: it sometimes coaches people who just wanted a straight answer.

Why it matters: most AI labs are racing to automate work away. This is a serious attempt at an AI that makes groups of people work better together, and it's honest about where it falls short.

TRY IT → through the humans& Agents API, with no waitlist. It costs about A$1.43 per million tokens in and A$7.43 per million out (US$1.00 and US$5.20), or about A$0.29 per million for cached input. It's for developers for now; humans& says an app built around it is coming soon. humans&: Introducing Ampersand · Pricing and limits

Also shipped

Skip this: paying per word for a voice. Between Phonon running free on your own device and Mistral's prices, try a cheap option on your real script before you sign up for an expensive voice plan.

Five to Try is curated with AI help and checked by a human. Hit reply and tell us what you tried.

Know someone who'd like this? Forward it on. They can sign up free at https://fivetotry.beehiiv.com/subscribe