All About AI

Claude Fable 5 Agentic AI Trading: First Tests Look VERY Strong

By · · 6 min read

Anthropic just dropped Claude Fable 5 and I ran it overnight on a Polymarket 5-minute up-and-down bot. Result so far: +$41 at the 10-hour mark with a 71% win rate, then +$82 over the first 24 hours, almost $100 profit by the time I sat down to record. But what surprised me wasn't the number. It was the strategy Fable 5 picked on its own — specifically a "deep long-shot fading a jump" branch I haven't seen any model construct before. Plus a self-monitoring cron loop that ran adjustments while I slept.

Watch the video:

The setup — and the lucky timing

I'd collected 24 hours of Polymarket 5-minute up-and-down market data just a few hours before Anthropic released Fable 5. Not planned — just lucky. So when the model dropped, I had a fresh, untouched dataset sitting in the same shape I always use for new-model tests:

  1. Point the model at the raw data
  2. Tell it to analyze, design a +EV strategy, build the bot, execute
  3. Watch what kind of strategy emerges

This is the same harness I use for every new release. Same harness I used in the Opus 4.8 first test and the head-to-head from Codex vs Claude on Polymarket. The data side comes from the five-source data pipeline I just covered — collect everything into one master file, point the agent at it.

The prompt I used was basically: "From the 5-minute up-and-down market data, analyze, design a +EV strategy, write the tests you'll need, think hard, use your 100x financial genius brain." (The 100x line is for fun — Fable didn't bite, just went straight into the work.)

The strategy Fable 5 came up with

After spending a substantial chunk of tokens on per-snapshot testing — which is something I haven't seen prior models do at this granularity — Fable 5 landed on a recommended live setup:

The fee accounting in the formula is the part I want to flag — calculating forward profits without subtracting venue fees is one of the most common mistakes in retail-built bots. Fable baked it into the signal definition, not as an afterthought.

The "deep long-shot fading a jump" branch

This is the part that genuinely surprised me. Inside the strategy report, Fable 5 included a sub-branch that takes positions the model itself thinks will probably lose. From the model's own explanation:

This is a trade that would probably lose. The model only gives it 22.7%. That is not a malfunction or a bad entry. It's +EV because the $59 wins dwarf the $13 loss. Break-even is 18%. Model says 23%. Hence the +EV, ~4.5 cent edge.

That's an asymmetric long-shot fade — entering at low prices where the win probability is small but the payoff ratio more than compensates. I've seen models avoid trades they think will lose. I've seen models size up trades they think will win. I haven't seen a model voluntarily split its strategy into "conservative core" + "intentional long-shot fades" with the math justifying the latter on payoff geometry alone. That's a real planning move, not a prompt artifact.

The cron self-monitor — trade while you sleep

Here's the part that made the overnight run actually work. Three hours in, I told Fable:

We've been trading for 3 hours. Analyze trades so far, make adjustments if you see any to improve EV. If not, decide what to do.

It re-tuned slightly and the win rate ticked up. Then I needed to sleep, but I wanted the bot to keep running. So I asked for a self-monitor:

Can you set up a monitor that wakes every 2 hours, checks the trades, makes adjustments if needed, and restarts the bot?

Fable used Claude Code's cron support to schedule itself. Every two hours overnight: wake up, run a health check, analyze every trade since the last check, verify the strategy formula still has edge, adjust on strong evidence, restart if warranted. Five scheduled reviews ran while I slept. All five came back "bot healthy, no changes warranted." The fifth one specifically: "session is now very positive, no changes."

This is the cleaner version of the heartbeat split-agent setup from earlier. There I built it manually with sub-agents. Fable just used Claude Code's built-in cron and self-scheduled. Less code, same shape.

The interesting positions it took

Scrolling through the trade log, the span of position sizes Fable picked was wider than I'd seen on prior models:

One honest admission: the early long-shot entries were too aggressive. Fable itself flagged later that if it had tightened the cheap-side entries, the +$41 ledger would have been closer to +$120. It corrected through the cron adjustments and hasn't taken those super-low entries since. That kind of self-correction inside a live run is the property I actually care about — model that's wrong, that knows it, and updates on its own evidence.

Fable 5 burns tokens — use model switching

The one real caveat: Fable 5 is token-hungry. Running everything on Fable will drain a Pro/Max account fast. The fix is model switching inside the same Claude Code session:

I ran the trade-strategy explainer (Fable built an interactive HTML site showing the formula, the Binance-leads-Polymarket-lags edge, the entry conditions, and the trade ledger broken out into "winning trades" vs "bad trades" sections) on Fable extra-high, but pulled all the directory lookups and log greps onto Sonnet. Roughly 80% of the operations are sub-tasks Sonnet handles fine.

What this tells me about Fable 5 for agentic trading

Three things stand out after the first 24 hours:

  1. Per-snapshot testing depth. Fable ran tests on individual data snapshots, not just the aggregate. The strategies it produces are visibly informed by more granular evidence than what prior models build from the same dataset.
  2. Asymmetric strategy composition. The long-shot fade branch is the first time I've seen a model construct a deliberately-losing-most-of-the-time sub-strategy on its own and justify it with payoff math. This is real planning, not template-matching.
  3. Self-correction in live runs. The cron-driven adjustment loop actually fixed its own bad early entries. Most models in this space don't update mid-run — they keep doing the same wrong thing until you stop them.

The negatives are real too. Token cost is high. The initial long-shot sizing was too aggressive and would have cost real money if I hadn't given it the 3-hour review prompt. Hold-to-resolution with no stop loss is a strong choice that works on Polymarket's microstructure but won't transfer cleanly to perp DEXes — anyone trying to repurpose this onto Hyperliquid needs to redesign the exit logic.

What's next

I'm leaving the bot running and will keep posting updates in the AI_automata Discord as the sample size grows. 15 hours and ~$100 isn't a sample, it's a vibe. The real question is whether the win rate and the long-shot fades hold over 200+ trades. Next video I'll likely run Fable 5 head-to-head against the previous best — Opus 4.7 and Codex 5.5 — on the same dataset to isolate how much of this is the model vs how much was just a good 24 hours of market regime.

Resources