Opus 5.5 Wrote a Tycoon Game From Scratch That Trades Real Money: the Day Trader Tycoon Experiment

You open a position in the market and the same position appears in your real exchange account seconds later; the only difference is that one sits on a game screen and the other in a live wallet. All About AI host Chris does exactly this in Day Trader Tycoon, a game written from scratch with Opus 5.5 : an order issued from the game menu travels over a websocket connection into his Hyperliquid account, and the confirmation lands in real time. He sums up the seriousness in one sentence, saying he connected one of his accounts to the game and stressing that it is no joke.

Retro brief and a one person game studio

The birth certificate of the game is a single prompt: build a day trader tycoon game in the style of RollerCoaster Tycoon, channeling the spirit of OpenRCT2 . The host points to the open-source remake of the childhood classic he grew up with; the OpenRCT2 community has likewise rebuilt the original RollerCoaster Tycoon 2 as an open-source rewrite, now at version 0.5.4 and distributed free of charge through openrct2.org. The model takes this brief and builds an entire game complete with its economy; in the host's words, there is no longer any need for insanely detailed instructions.

The critical distinction is this: the game is no browser toy but a Mac application written end to end in C++ , with a framework on Apple Metal underneath the graphics. The model produced not just the game logic but every file of a compiled desktop application. That claim may sound inflated, yet the picture on the Anthropic side supports it, since the company announced with Opus 4.5 that the model was breaking ground on real-world software engineering tests. Reuters covered the same launch as a leap in coding and agent abilities, while Anthropic introduced that version as its best coding model to date.

So where did that power go one version later? According to the official documentation on platform.claude.com, Claude Opus 5.5 arrived on September 22, 2026 as the first member of the new 5.5 family, designed for long-running agentic coding. The docs state the model performs at the level of Fable 5.1, keeps adaptive thinking always on, and is priced at 4 dollars input and 20 dollars output per million tokens. The experience in the video matches that description exactly: the model shuts the game down, takes screenshots, tests the feature, and reports back when done.

Three game modes and day one of trader life

The game offers three different starts: real markets, simulated markets, and offline snapshots. Real markets mode pulls live prices into an offline simulated environment; simulated markets feel livelier; snapshots exist for playing without internet. Prices from Yahoo Finance, Hyperliquid, Kalshi, and Polymarket merge into a single pool. In this section the host first deletes his old saves and starts clean.

The character creation screen gives away the humor of the game: a name is typed, a look is randomized, and a trading style is picked, with the host torn between gambler and whale before choosing whale. Starting capital is tight, the venue is a laptop in a basement, and the money counter on screen tells the whole story. The first trade brings a level-up, with points flowing into stats like charisma while the title line starts at intern. The loop is the classic tycoon formula: start small, win, grow.

Construction is the key to growth. A multi monitor rig is bought and placed in the basement, unlocking fresh trading options. An equities desk is desired but cannot be built without an office, so every market access carries a precondition. Prediction markets need their own prediction desk; a question like what level crude oil reaches in October stays locked until that desk exists. A client system kicks in later, drawing outside investors into the growing fund.

Leverage, news terminal, and the main save

The trading screen is designed with brokerage seriousness: with a 5 thousand bankroll, 5x leverage is selected, a 2 thousand position grows to 10 thousand in size, and a 5 percent stop-loss with a 20 percent target goes in. The portfolio screen tracks metrics like maximum drawdown and funding return. In later saves, with Bitcoin at the 81 level, 40x leverage unlocks and a 2 million position becomes theoretically possible. The host closes a small profit and moves on; the numbers may be play money, but the decision interface is the real thing.

The beating heart of the game is the ever-scrolling news terminal : a rumor of crude falling to 89 dollars, rising stocks, a sliding US 10-year yield, and gold headlines keep dropping. Every story carries a probability next to a market price; with deflation priced at 76 percent against a market of 77, no edge remains. Game speeds of three and five melt waiting times away. Sellable clutter like a pizza box and a bean bag completes the set.

Loading the main save changes the scale: a large office with a crypto desk, a predictions desk, a macro desk, an equities desk, a perp pit, and a risk desk side by side. The staff screen holds a risk manager and traders, with the quant trader credited 2 thousand dollars earned. The synergy system pays small but tangible bonuses, with a quant-heavy roster adding basis points of edge and rebalancing twice as often. House style choice spreads different bonuses across setups, whether quant shop, prop desk, or research house.

Rivalry lives on the leaderboard. Competing funds line up, Granite Street on top, with sizes of 17 and 16 million mentioned and a 12 percent monthly return counting as strong. The host fund drifts between fourth and fifth while a rival must have booked a loss on a bad bet. A hiring screen and prestige spending like a private island decorate the race. A challenges panel pays per goal: 50 thousand for a millionaire title, a separate prize for a won prediction bet.

Real orders, prediction bets, and a news reading agent

Back in live mode things turn serious: a binary booth is built and placed on the plaza, opening a five-minute up-or-down market. The host confirms a 1 dollar down bet and the position genuinely appears on Polymarket, cashable from the portfolio screen. Windows of ten, fifteen minutes, and one hour wait in the list. Polymarket, described as the largest prediction market on earth, offers market data, order placement, and SDK tooling in its developer docs on docs.polymarket.com; the integration in the video is that API surface moved inside a game.

The same account-merging logic works on the equity side: Nvidia, Google, and Tesla shares trade alongside Brent crude and the S and P 500 through perpetual contracts on Hyperliquid. According to the official Hyperliquid docs hosted on GitBook, perpetual data comes from the info endpoint at api.hyperliquid.xyz, with perpDexs and meta calls returning universe, margin tables, and prices. The host setup mirrors exactly that architecture: a raw websocket connection, a local server in between, and one screen uniting Hyperliquid, Polymarket, and brokerage accounts alike.

The brightest segment of the video is the feature order: since early game funds cannot hire anyone, the host requests an AI agent system that reads the news and trades automatically. The terms are written crisply: a level 2 requirement, a 3 thousand price tag, an allocated budget, trading only on unlocked desks, and learning over time. The model closes the game, moves to the dev environment, tests with screenshots, and reports when finished. Balancing runs eight agents over thirty market days, with thousands of simulated days converging on a 52 percent average win rate.

In the live trial the agent genuinely reads the feed: it catches a headline of 20 thousand Bitcoin pulled from exchanges within an hour, computes its own 70 percent odds against a 64 market price, and opens a long at 5x leverage. The position moves 10 dollars into profit and lands in the portfolio while the account sits 3 percent up. A thinking log stays readable on screen with an adjustable edge value. The host says the feature took about twenty minutes to add, tying open-source or Steam release plans to a thousand likes. From the day-night cycle to every C++ file, the whole thing is the work of one person plus a model.