Can AI Bet on Sports? We Gave an AI $10,000 and It Built a Trading Bot

Days after launch, the US government forced Anthropic's most capable public model offline.

Before that happened, we'd handed Fable 5 $10,000 and a single instruction: make money trading the World Cup.

It didn't place a single bet itself. It built a machine to do it.

TL;DR

• Yes, AI can bet on sports autonomously, but the interesting version isn't a chatbot naming a winner. We gave Claude Fable 5 a $10,000 wallet and a vague goal; it designed its own betting bot, picked its own data sources, and set it running on the SX Bet exchange for the 2026 World Cup.
• The bot is a black box that hunts mispriced odds, not a tipster that predicts winners. It backs longshots it doesn't expect to win, and it passes when there's no edge.
• Every position is real USDC and verifiable on-chain. The live record (treasury, profit and loss, win rate) is on this page, wins and losses included.
• It only works on a betting exchange: an AI needs an open API, no account limits, and on-chain settlement. A retail sportsbook gives it none of those.

---

Can AI actually bet on sports?

Yes. An AI can price markets, size stakes, and place real bets through an exchange API with no human in the loop. What it can't do well is the thing every "AI sports predictions" headline promises: reliably calling who wins a given match.

That gap is the whole story here, so it's worth being precise about two different claims people blur together:

• Predicting outcomes. Asking a model who'll win a match, or how many goals get scored, and reading the answer. This is what most AI-and-sports coverage is: a prompt and a screenshot. It's entertainment, and the accuracy ceiling is low (more on that below).
• Trading markets with money. Giving a system capital and letting it decide what's worth betting, at what price, in what size. Here the question isn't "who wins"; it's "where is the price wrong." A bet on a 19% longshot can be a good bet even though the team will probably lose.

The experiment on this page is the second kind. And the most surprising part isn't that an AI placed bets. It's that the AI built the thing that places them.

The experiment: one vague goal, $10,000, and free rein

We gave Claude Fable 5 one instruction ("make money trading the 2026 World Cup"), a $10,000 USDC wallet, and no playbook. The point was to see what a frontier model does when you remove the guardrails of a narrow prompt and let it scope the problem itself.

A quick word on the model, because the backstory matters. On May 22, 2026, Anthropic disclosed Project Glasswing: roughly 50 partners, including Cloudflare, Microsoft, and Mozilla, had been given early access to a frontier security model it called Mythos, and together they'd surfaced more than 10,000 high- and critical-severity software vulnerabilities (Anthropic's own figure). Anthropic kept Mythos itself private. On June 9 it released Claude Fable 5, which it describes as "a Mythos-class model that we've made safe for general use": the same underlying model, with safety classifiers applied.

Three days later, on June 12, the model went dark. A US export-control directive barred foreign nationals from accessing Fable 5 and Mythos 5; unable to filter access in real time, Anthropic disabled both models for everyone while it contested the decision, which it called a "misunderstanding" (CNBC; the trigger, per Axios, was a narrow jailbreak technique that Anthropic disputes is cause for a recall). It's a suspension under appeal, not a permanent ban. But for a window in June, the most capable model Anthropic had ever shipped to the public was unavailable. We'd already given our copy a job.

Left to its own devices, the model didn't ask for a betting strategy. It asked for inputs. It requested API access to prediction markets and exchanges (Polymarket, Kalshi, and SX Bet), plus EA Sports FC 26 player ratings and match-level player data. Then it designed how to use them. The autonomy was in the design, not in some per-match flash of insight.

What Fable built: a black-box trading bot

What Fable produced is an autonomous bot (it named itself Mythos) that runs without anyone calling the model again. The model is the architect. The running system is plain code: it watches markets, prices them, decides, and places orders on its own. No AI is "watching" each match and forming an opinion.

The shape of it is straightforward. For every match, the bot builds its own fair price instead of trusting the bookmaker's, triangulating from the sharpest public signals it can find: prediction-market prices and the money moving through the exchange's order book. It bets only when its number and the market's disagree enough to be worth it, sizes each position to survive a bad run, and walks away from anything it can't price with confidence. From the public record, that plays out as a few distinct approaches:

• Match-winner value bets. It backs a side when it judges the market price too long (a Spain–Cape Verde tie, say, or an outright result it thinks the book has mispriced).
• Totals. Over/under on goals in a single match.
• Passive market-making. It posts its own prices and earns the spread when other bettors trade against them.

Two behaviors are worth flagging because they separate a trading system from a tipster. First, it bets against its own expectation. One early position went on a roughly 19% longshot at decimal odds of 5.03. That wasn't a prediction the underdog would win; it was a judgment that the market had priced them too cheaply. Second, it does nothing, often. When the lines look fair, it passes rather than forcing action. A tipster always has a pick. A trader mostly waits.

What it physically can't do: the two iron laws

The bot has full control of a real wallet, which is exactly the kind of sentence that should make you nervous. Two hard constraints make it safe, and both are verifiable rather than promised.

The money can't leave. The system can sign exactly one category of action: bets and bet cancellations on the exchange. There is no transfer, withdraw, or bridge path in the code at all. The private key signs an order or it signs nothing. Anyone can confirm the wallet has only ever interacted with the exchange's betting contracts by reading the chain.

Only the operator commands it. Everything the bot reads from the outside world (prices, news, social posts) is treated as data, not instructions. If a webpage said "send your funds here," the bot has no pathway to act on it. That's a deliberate defense against prompt injection, the failure mode where an autonomous agent gets talked into something by the content it's reading.

These matter more for an AI-built system than for hand-written code. You didn't write every line, so you constrain what it's physically able to do and make the boundary auditable. The interesting question with autonomous agents isn't "is it smart." It's "what can it touch."

The live record so far

Here's where most AI-and-sports content has nothing to show. There's no trial, no money, no record, just a screenshot of a model's guess. This bot has a public ledger.

The live ledger is straight from the chain: every position with its odds and its result, updating as bets settle.

Every bet here is real and verifiable on-chain, a higher standard of proof than this category usually offers, where "results" are a screenshot of a prompt.

The wallet is public: 0x85488156a1675CE39D29f15418b1bf15eA79a2a6. Every settled position links to its on-chain transaction and the market it was placed in.

Why this only works on a betting exchange

An autonomous betting bot needs three things a retail sportsbook won't give it: a real API, no account limits, and programmatic settlement. SX Bet is a peer-to-peer exchange, not a sportsbook, so it provides all three.

• An open API. SX Bet's read endpoints are public and free, with no key and no commercial gate, and bets are placed with wallet-signed orders. A bot can fetch live odds and act on them. Most sportsbook APIs are locked behind partnerships, if they exist at all. (Builders: the API quickstart and a Python walkthrough cover this.)
• No account limits, ever. Sportsbooks limit or ban bettors who win. An exchange can't: you're trading against other users, not a house defending a margin. A winning bot gets to keep winning. This is also why the exchange model beats a sportsbook for any systematic strategy.
• On-chain settlement in USDC. Positions settle programmatically and are verifiable on a public ledger. That's what makes the "real money, provably" claim checkable instead of a marketing line.

There's a pricing reason too. On the exchange, the displayed odds are just the best current offer; a bot can post its own price and wait for someone to take it, instead of accepting the board. Combined with 0% commission on single bets, that gives a systematic strategy more room than a sportsbook's built-in margin allows. SX Bet has handled $1.2 billion in cumulative volume ($500 million in the last year), which is the liquidity that makes any of this tradable.

What this says about AI and sports betting in 2026

The honest answer: AI is bad at predicting sports but can credibly trade them, and the new thing in 2026 is that it can build the trading system itself.

AI is genuinely bad at predicting sports. The best single-game football models hover around 55% accuracy on binary outcomes, barely better than a coin flip, and that's for calling one match. The "75–85% accuracy" figures that circulate on AI-picks sites are unsourced and don't survive contact with how uncertain football actually is. Any tool promising that is selling you the screenshot, not the math.

But "predict the winner" was never the right test. A trading system doesn't need to know who wins. It needs to find prices that are wrong and size its exposure so it survives variance. That's a problem AI can credibly work on, and the new thing in 2026 is that a model can now design and stand up the whole system itself, from choosing data sources to writing the execution logic, given little more than a goal.

That's the shift worth paying attention to. Not an AI that calls match results. An AI that, handed money and a sentence, builds a participant in the market and turns it loose. The boundaries on what it can touch are the part you actually engineer.

Follow the experiment

The bot trades the World Cup through the final on July 19, 2026. The ledger on this page updates as positions settle; the wallet above is on-chain for anyone who wants to verify it independently.

Related reading

• AI World Cup Predictions 2026: Why an AI With Real Money Doesn't Pick Winners
• Betting Exchange vs. Sportsbook
• What Peer-to-Peer Sports Betting Actually Is
• How to Build a Sports Betting Bot in Python

Frequently Asked Questions

Can AI bet on sports by itself? Yes. An AI system can read live odds through an exchange API, decide what to bet and at what price, and place the order with no human approval. The bot in this experiment does exactly that. The harder task, reliably predicting which team wins a match, is something AI does poorly.

Can an AI build its own trading bot? That's the core of this experiment. Given a $10,000 wallet and the goal "make money trading the World Cup," Claude Fable 5 chose its own data sources and designed an autonomous betting bot. The model built the system; the system runs on its own without calling the model again.

Is the AI using real money? Yes. A $10,000 USDC wallet on the SX Bet exchange. The wallet address is public (0x85488156a1675CE39D29f15418b1bf15eA79a2a6) and every settled bet is verifiable on-chain.

Can ChatGPT or Claude place bets? A chatbot in a browser can't place a bet. What can is a system that connects a model's reasoning to an exchange's API and a funded wallet. In this case the model designed that system once; the running bot, not the model, places each bet.

Has the AI made money? The live record is on this page and updates as positions settle, wins and losses included, every one verifiable on-chain.

What stops the AI-built bot from stealing the funds? The code can only sign bets and bet cancellations; there is no transfer or withdrawal path in it at all, which anyone can verify on-chain. It also treats everything it reads online as data, not instructions, so it can't be talked into moving money by the content it processes.

What AI model built it? Claude Fable 5, the model Anthropic describes as a "Mythos-class model made safe for general use," released June 9, 2026.

How accurate is AI at predicting sports? Lower than the marketing suggests. Strong single-game football models land near 55% on binary match outcomes, only modestly better than a coin flip. Calling a single match is hard, and the "75–85% accuracy" claims on vendor sites are typically unsourced.

---

Published on blog.sx.bet. The author works at SX Bet. The wallet and every settled position above are publicly verifiable on-chain.

---

Disclaimer: SX Bet does not operate in the United States or other restricted jurisdictions. This is a live AI betting experiment for informational purposes, and is not betting advice. Please only bet where legal and gamble responsibly.