G'day, and happy Wednesday. Five releases from the last day and a half, why each one matters, and how to try it. OpenAI had a big Tuesday morning (our time): a cheaper, faster model built to click around your screen for you, and a claim that its AI has knocked over a famous computer science assumption. In between, there's a free AI helper that runs entirely on your own computer, a US$5 app that tells you which coding agent is cooking your Mac, and Minecraft's new edition for Switch 2. About a five-minute read.

Image: OpenAI

OpenAI has released GPT-6.1 Terra, a fast, lower-cost model for everyday office work. Its party trick is "computer use": it looks at the screen, works out what's there, then clicks and types across your browser, desktop and apps. OpenAI's examples include copying invoice details into accounting software, filling in forms and turning technical drawings into 3D models. On OpenAI's own tests it scores 15 points higher than the old GPT-5.6 Terra on a set of real-world computer tasks at about a third of the cost. It also makes far fewer factual mistakes: on a set of trick questions designed to catch errors, the error rate fell from 53.6% to 16.7%. It's better at coding too, though OpenAI says its pricier GPT-6.1 Sol and GPT-6 Astra are still the pick for the hardest jobs. API prices are half those of GPT-5.6 Terra.

Why it matters: AI that can operate ordinary software, rather than just chat about it, is where the real time savings are for small businesses. This makes it a lot cheaper to try. All the benchmark numbers are OpenAI's own.

TRY IT → in ChatGPT Work (desktop app or web) and Codex on paid ChatGPT plans, from Plus up. Pick it from the model list; it isn't in the regular Chat mode. For developers, the API costs US$1 (about A$1.42) per million input tokens and US$5 (about A$7.12) per million output tokens. OpenAI: Introducing GPT-6.1 Terra · ChatGPT: what's new · API docs

Image: Frigade / Hob (GitHub)

The Claude and ChatGPT desktop apps can now work on your files: read a folder, edit a spreadsheet, run some code, draw a chart. Hob does the same kind of work, except the AI runs on your own computer and your chats and files never leave it. It's a free, open-source app from the software company Frigade. On first launch it checks your hardware and downloads the best open model it can run well (Google's Gemma 4 on most computers, OpenAI's gpt-oss on big workstations), so there's no fiddly setup. It can read PDFs (including scans), look at photos and screenshots, take dictation, search the web and run code in a locked-down sandbox. It asks before changing anything in your folders, and every task can be undone. Turn web access off and it works fully offline.

Why it matters: for plenty of everyday jobs, a free model on your own laptop is now good enough, and nothing private gets sent to anyone's cloud.

TRY IT → free and open source (MIT licence), with no account and no subscription. For Apple Silicon Macs, Windows and Linux, with at least 8 GB of memory (16 GB is better for everyday tasks). It's brand new, so expect rough edges. Hob · GitHub: FrigadeHQ/hob · Show HN post

Image: WhoIsCooking (whoiscooking.site)

If you leave Claude Code, Codex or Cursor building and testing in the background, you'll know the sound of a MacBook's fans going flat out. Activity Monitor just shows you 80 mystery processes. WhoIsCooking, a small Mac app from developer Andrew Wilson, sits in the notch or menu bar, shows your chip temperature and puts each agent session (and everything it started) on one line, so you can see who's to blame. One click moves the busy work onto the Mac's cooler, slower "efficiency" cores, so the job keeps running instead of being killed. It also nudges you when an agent has been burning CPU for five minutes straight or left dev servers running. It recognises Claude Code, Codex CLI, Gemini CLI, Cursor, Aider, Copilot CLI, Ollama and LM Studio.

Why it matters: running several coding agents at once is now normal, and it's a small, honest fix for a real annoyance, without stopping work halfway through.

TRY IT → US$5 (about A$7.12) once, with no subscription and no account. Needs an Apple Silicon Mac (M1 to M5) on macOS 14 or later. It can't cool GPU-heavy work or Apple's own processes. WhoIsCooking · Product Hunt launch

Image: Mojang / Nintendo eShop (Australia)

Minecraft has a new edition built for the Nintendo Switch 2. The headline is "Vibrant Visuals", Minecraft's prettier lighting and shadows, which is switched on by default (though not in split-screen or some Marketplace worlds). The Switch-only Super Mario Mash-Up pack, with Peach's Castle, Delfino Plaza and Mario skins, has been updated to match. In Australia the boxed version came out yesterday, and the digital version unlocks today. If you already own Minecraft digitally on the original Switch, you can grab the Switch 2 Standard Edition free until the end of January 2027.

Why it matters: it's one of the most-played games in the world, the school holidays are coming, and for many families this is a free upgrade, not a new purchase.

TRY IT → A$45.95 for the digital Standard Edition on the Nintendo eShop, with Deluxe and Ultimate Collections that add extra packs and Minecoins. Free for existing digital owners of the Switch version until the end of January 2027 (it doesn't apply to cartridge copies). Boxed copies are at retailers, with prices that vary. Nintendo AU: Minecraft for Switch 2 · Nintendo AU: pre-order and upgrade details

Image: OpenAI (openai/math on GitHub), first page of the paper

For decades, computer scientists have assumed that some puzzles, like checking whether a long list of true-or-false rules can all be satisfied at once, can't be solved much faster than by trying nearly every combination. That assumption, the "Exponential Time Hypothesis", underpins hundreds of research papers. On Tuesday, OpenAI said one of its internal AI systems has found a faster method that disproves it, and that the proof has been checked in Lean, software that verifies maths step by step. Two of the researchers who first proposed the hypothesis, plus MIT's Ryan Williams, will lead an outside group to examine it. OpenAI also claims new results in maths (the Kakeya conjecture in every dimension) and physics, though the physics material isn't public yet.

Why it matters: if it holds up, an AI has overturned a core belief in its own field. That's a very big "if", and experts are only starting to check it, so treat it as a claim for now.

TRY IT → free to read. The 60-odd-page paper and the Lean proof are public on GitHub, and OpenAI's post explains the three results in plain-ish English. OpenAI: the announcement · GitHub: the ETH paper and Lean proof

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

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

Want your product in front of curious Aussie tech readers? Sponsor Five to Try at [email protected].