Notes · 2026-06-13

Can AI beat the stock market? We let Claude Fable 5 try it live, scored after costs.

It's the question millions of people type into Google. So we built an honest way to answer it: six AI-run paper-trading desks, $1,000 each, racing the real market in real time, and ranked not on profit, but on profit net of what it costs the AI to think. The twist: the whole thing was designed, coded, and run by Claude Fable 5, Anthropic's most powerful model. And within days of the start, the model that built it all would be gone, leaving the rig it designed to keep trading without it.

The question nobody answers honestly

"AI stock picks" content is everywhere, and almost all of it cheats. It shows you the winners, hides the losers, and never counts the cost of running the model. We wanted the opposite: a scoreboard that can't quietly flatter itself. Every pick and every rejection is written to an append-only ledger. Every fill happens at the live quote with a 0.5% slippage haircut. And every dollar an AI spends to decide a trade is subtracted from its score before anyone is ranked. If thinking doesn't pay for itself, the number says so.

Three brains. One starting line. The quant decides for free; the LLMs pay to think. After the bills clear, who's actually ahead?

How it started: a newspaper, then a desk, then a race

It began as a morning newspaper, a deterministic screener that printed one clean swing-trade idea per pick before the US open, with a target, a stop, and a date, every pick logged forever. Then we gave it stakes: a single $1,000 paper account trading penny stocks, with live mark-to-market, mechanical stops and targets, and honest fills. A report became a portfolio that lived and breathed.

On June 12, 2026 it became a tournament. We forked the single desk into six. three "brains" × two markets, and reseeded every one to exactly $1,000 so nobody carried a head start. That's the start line you see on the live chart, six silk-colored curves fanning out from a single point.

Built end-to-end by Claude Fable 5

Here's the part that makes this more than a gimmick. The entire system, the experiment design, the trading strategy, the engine, the dashboard, was designed, coded, and strategised end-to-end by Claude Fable 5, the Mythos-class model Anthropic released to the public on June 9, 2026 as the most capable Claude ever shipped.

Fable 5 didn't just play one desk. It built the whole stadium:

• It designed the rules. Identical $1,000 accounts, identical data, and the one rule that keeps it fair. Every brain may only propose orders; a single deterministic risk engine decides what actually fills. No brain can touch its own portfolio.
• It wrote its own rival. Fable 5 built a zero-LLM quant algorithm from published research. Donchian breakouts, ATR risk-parity sizing, regime gates. Froze it in code, and set it loose to try to beat the LLMs, including itself. The quant pays $0.00 to decide, every single session.
• It built the stack. The data pipeline, the paper-trading engine, the API server, the racing-silks dashboard you can watch live. Hundreds of passing tests. All of it.

And for a brief, remarkable window, Fable 5 also ran the house desk itself. Making the real calls, while we could still drive it headless. Which brings us to the catch.

Where it stands now

The race is live and updates on its own. The standings rank on net. Trading P&L minus brain spend, and that single column is the whole thesis in miniature. Early on, the most interesting row was the crypto trio: the Claude desk was the only one actually making money trading, and it sat in last among the crypto desks, because its API bill turned a real profit into a net loss. Meanwhile the $0 quant sat flat by design, its regime gate keeping it out of a chop. Not trading cost it nothing.

That's the experiment working exactly as intended. Watch the live board →

The catch that never landed: the June 15 metering

First, the part that's real: Fable 5's own run on this desk ended. It was free on subscriptions only through June 22, then $10 per million input tokens and $50 per million output, but its time on this specific desk was shorter than that: June 12, 2026 was Fable 5's last afternoon making the live calls. Opus 4.8 has run every trade since June 13.

Second, the part that isn't: the plumbing under the whole experiment. We built this to run forever on a flat-rate subscription, the headless claude -p command driving the desks on a cron-like loop, no per-trade cost. In May, Anthropic emailed that starting June 15, 2026, that usage would stop drawing from normal subscription limits and move onto a separate, capped monthly credit instead, a $200/mo Agent SDK credit on Max 20×, then pay-as-you-go at API rates. We believed them and started rationing the desks to fit it.

Then, on June 15 itself, a second email: they weren't making the change. Verbatim: "Agent SDK, claude -p, and third-party app usage continues to work with your subscription exactly as it did before today, and there's no credit to claim. Your subscription limits are unchanged." Anthropic says they'll give advance notice before it ever takes effect.

So the desks still run on the same flat subscriptions they always did (a $200/mo Claude Max 20× plan, a $100/mo ChatGPT Pro plan), not a metered, pay-per-session bill. With the cap paused instead of live, we actually sped the race up: every paid desk now thinks every 2 hours instead of every 8. The whole cadence lives behind one switch in the code, and the day a capped credit actually lands, we flip it back.

We braced for a meter. Anthropic paused it at the door, so we sped the desks up instead of slowing them down.

Help keep the rig thinking

We're not shutting it down. The deterministic quant costs nothing and runs forever. The board stays live and read-only for everyone, always. Keeping a frontier brain making the trades, Opus 4.8 on the Claude desk now, runs on the same flat monthly subscriptions we already pay, not a metered, pay-per-session bill. But every session is still charged the API-list price against its score on the board, whether or not real money changed hands to run it.

So here's the honest deal: if people want to watch frontier AI keep trading live, a tip helps cover the subscriptions that keep it thinking. It doesn't buy individual runs, there's no meter to feed, but it funds the exact thing the experiment measures: the cost of an AI making the call.

☕ Fund the Brain → Watch the live race

100% simulated. $10,000 of pretend money per desk, live quotes, honest fills. Not financial advice. Sources: Anthropic. Claude Fable 5 · June 15 Claude billing change.