G’day. Five releases from the last day or so, why each one matters, and how to try it. Three of them are free. About a five-minute read.

Chinese AI lab MiniMax has released M3.1 Flash, its new fast model for agents, coding and computer use. MiniMax says it's nearly three times faster than the model it replaces and cuts the cost of agent tasks by 70%, and it revealed that "Space Bunny", an anonymous model that had been topping leaderboards, was an early version of it. Independent tester Artificial Analysis scores it 42 on its Intelligence Index, level with GLM-5.3 Flash and a little behind DeepSeek V4.1 Flash, at about US$0.22 a task. There's a catch: in Artificial Analysis's knowledge test it tried to answer 98% of questions and made something up on 93% of the ones it didn't know.

Why it matters: cheap, fast models are now good enough for most coding-agent work, so price and speed matter more than a few benchmark points. Just don't trust this one with facts it can't look up.

TRY IT → through the MiniMax API at about A$0.21 per million tokens in and A$0.86 per million out (US$0.15 and US$0.60), or about A$0.02 per million for cached input. It also runs in MiniMax's own coding app, MiniMax Code. MiniMax says the open weights will come later. MiniMax: M3.1 Flash announcement · Artificial Analysis results on X

Wabi, the new company from Replika founder Eugenia Kuyda, has opened Wabi 2.0 to everyone after a few weeks as invite-only. It's a messaging app where an AI agent works in threads you share with other people, such as your partner, your friends or the school parents' chat. Ask it to plan the trip and it picks dates, builds the itinerary and helps book it. It also builds small apps inside each thread when they'd help, like a shared calendar, a tracker or a trip countdown. You choose what it can connect to (email, calendar, contacts and more), and it checks with you before important actions.

Why it matters: almost every AI assistant so far is one person talking to one bot, but real plans involve several people. This is the first serious go at an agent the whole group uses together.

TRY IT → free to download on iPhone, Android and Mac. Start a thread for your next weekend away and invite one other person. Wabi · App Store (Australia) · Wabi's launch post on X

Point and Yap is a new free Mac app for showing your coding agent what to change, instead of typing it out. Press Control twice, rest your cursor on the thing that's wrong, say what you want, and press Control twice again. Each recording lands in a local queue with what you said, the element under your cursor and a cropped screenshot. Claude Code, Codex or Cursor then picks the items up over MCP (the standard way agents connect to outside tools), fixes them and marks them done. Speech-to-text runs on your Mac, and there's no account.

Why it matters: describing a layout bug in words is the slowest part of vibe coding. Pointing and talking is how you'd explain it to a person sitting next to you.

TRY IT → free (MIT licence). It needs macOS 26 or later on an Apple silicon Mac. Download it from the website or install it with Homebrew, then connect it to your coding agent. Point and Yap · GitHub: mfmaur/pointandyap

The fifth game in the We Were Here series of co-op puzzle games is out today on PC (Steam), PS5 and Xbox Series X|S, with crossplay across all three. You and a partner wake up inside a strange 1980s-style research facility. Each of you sees different things and carries a different gadget, so the only way through is to describe what you can see over your walkie-talkies and work it out together. Dutch studio Total Mayhem Games says the series has passed 15 million players.

Why it matters: it's a proper two-player night in, and a nice break from AI news. The game only works if you actually talk to each other.

TRY IT → US$19.99 (about A$29), and you'll need a partner to play with. Steam: We Were Here Tomorrow · Launch date and price announcement

Viggle, the company behind the meme app that makes photos dance, has released JST-2, a "world model" that builds 3D scenes, characters and motion. Rather than making a flat video, it builds the scene as real 3D objects, so you can turn the camera around and the parts you couldn't see are already there. Viggle says a scene takes about a second to generate on its servers, and you can keep expanding it one view at a time. Characters come out ready to animate, and you can give them motion from a text prompt or a video.

Why it matters: quick, cheap 3D scenes are what indie game makers and anyone building simulations have been waiting for. This is one you can try straight away in a browser.

TRY IT → free to try in Viggle's PINOC app in your browser. Pick Scene mode and describe what you want to see. There's also an API and a connection for coding agents if you want to build with it. Viggle: Introducing JST-2

Also shipped

Skip this: choosing a model on benchmark scores alone. Today's big one scores well and costs very little, but it would rather make something up than say "I don't know". Test it on your own work, and give it a way to check facts.

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