G'day, and happy Melbourne Cup Day to those who get the day off. Five releases from the last couple of days, why each one matters, and how to try it. Meituan, the Chinese food-delivery giant, has released an AI model it says scores between two recent versions of Anthropic's top Claude model, and DeepSeek has shared a coding agent that waits in your team chat for jobs. There's also a bookshelf of the books Hacker News readers can't stop recommending, a puzzle game where every word on screen is in an alien script, and a new test of AI on expert work where the best model scores 12%. About a five-minute read, so you'll be done well before the race.

Image: LongCat (screenshot)

Meituan is best known in China for food delivery, but its AI team has just released LongCat-2.5, a huge model built for long jobs like writing whole software projects. It has 1.6 trillion "parameters" (the internal settings an AI learns), though only about 48 billion are used for each word it writes, which keeps it faster and cheaper to run. It understands images as well as text, which Meituan says helps it build interactive web pages and 3D showcase pages, and help with design. It also plugs into popular coding tools like Claude Code and Codex. Meituan says it trained the model entirely on Chinese-made chips, and that its test scores land between Anthropic's Claude Opus 4.8 and Claude Opus 5.

Why it matters: a top-tier AI trained on home-grown chips shows China's tech giants can keep up without American hardware, and more competition helps keep AI prices down for everyone. The test scores are Meituan's own.

TRY IT → Meituan says you can try it at longcat.ai. Developers pay US$0.30 (about A$0.43) per million "tokens" (small chunks of text) going in and US$1.20 (about A$1.71) per million coming out, a "limited-time" discounted price on LongCat's API. We couldn't check whether signing up works from Australia. LongCat: Introducing LongCat-2.5 · LongCat-2.5 pricing

Image: DeepSeek on GitHub (screenshot)

Most AI coding helpers sit and wait for you to type a request. DeepSeek, the Chinese lab behind the low-cost DeepSeek models, has shared a prototype "super-agent" that works more like a teammate. You connect it to one GitHub project (and, if your team uses it, the Feishu work chat app). When someone who's allowed asks it for something, it opens a job, works on it in its own space and replies in the same conversation, for example with finished code ready to review. A web dashboard shows anything that needs a person. DeepSeek's rule for it: "take as little human attention as possible, and give as much value as possible", so it only acts when asked or when a rule your team wrote allows it.

Why it matters: the next step for coding AI is agents that run all day and take work off a whole team, not just one person at a keyboard. DeepSeek has published its version for anyone to study and copy.

TRY IT → free and open source (MIT licence), and it shot up GitHub's charts with more than 10,000 stars. It's for developers: it runs on DeepSeek Harness (DeepSeek's free, open-source agent app) and needs an AI model provider, which you pay for separately. DeepSeek calls it a research prototype, "not an official DeepSeek product", and says to review and adapt it before using it, so try it on a spare project first. GitHub: dsh-superagent-prototype · DeepSeek Harness

Image: Hacker Books (screenshot)

Last issue we featured HackerTube, the slop-free YouTube Andrej Karpathy built by talking to an AI. Within a day, someone posting as "roansh" did the same for books. HackerBooks gathers the books that people on Hacker News (a popular tech discussion site) have mentioned in comments since 2007: 15,766 books from more than 225,000 comments. Each book links back to the conversations that mentioned it, and you can browse by topic or decade or see what's rising. Because mentioning a book isn't the same as liking it, an AI (Claude) labelled every mention as a recommendation, neutral or a complaint. Its maker says Claude wrote "almost all of it, including the data pipeline and the site", while they "mostly steered and made calls".

Why it matters: it's a lovely way to find your next read, and proof that a good idea can spread in a day once AI does the heavy lifting.

TRY IT → free, no sign-up, in any web browser. Book pages link to Bookshop.org and Amazon, and the site earns a small commission if you buy through them (it says so on its About page). The data is free to download. HackerBooks · Hacker News: Show HN: HackerBooks

Image: Artless Games (Steam)

In 2058, Earth starts receiving strange symbols from space, and it's your job to work out what they mean. Message From Aliens is a sci-fi puzzle game in which the whole interface is in an alien script of more than 1,000 symbols, and the only human words you'll see are your own notes. It eases you in gently: the first chapter is mostly alien maths, so you start by cracking their numbers, and by the end you're reading alien poetry and alien jokes. There's an in-game alien dictionary, and you don't have to translate everything to reach the ending. It's inspired by the books Contact and Story of Your Life, the story behind the film Arrival.

Why it matters: it's the language-decoding fantasy of Arrival turned into a game, from independent developer Artless Games, and it launched on Monday, Australian time.

TRY IT → out now on PC (Steam) for A$12.85, Windows only, with a free demo. The developer says a Mac version is coming "in a few weeks", and, charmingly, apologised for forgetting to switch on a launch discount. The developer says it's not recommended for players under 15. Steam: Message From Aliens

Image: SemiAnalysis GDPx (screenshot)

AI companies love to boast about test scores, but many popular tests are now too easy. Research firm SemiAnalysis has launched GDPx, 206 hard tasks across 11 valuable fields including chip design, cybersecurity, biology, finance, law and share trading, built with 14 specialist partner companies. Each AI gets up to eight hours for most tasks, usually working in its maker's own coding tools. The best result is Anthropic's Claude Opus 5.5 at just 12%, ahead of Google's Gemini 4 Carbon (10.9%) and OpenAI's GPT-6 Astra (9.6%). SemiAnalysis says its finance and law tasks are "so challenging that basically every model scores exactly 0".

Why it matters: it's a useful reality check. Today's best AI is impressive, but it's still a long way from doing an expert's whole job on its own. Some models also lose marks because their safety filters refuse risky biology or hacking tasks.

TRY IT → free to browse the leaderboard, compare models by field and see the average cost per task (about US$51, or A$73, for Claude Opus 5.5). SemiAnalysis's full write-up is partly for paid subscribers. GDPx leaderboard · SemiAnalysis: GDPx

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].