How Our Claude Bot Made $200K Profit This Month
Source: http://x.com/i/article/2040396788968955904
And why Claude has nothing to do with why it works.
Yeah we use Claude. Claude wrote the code. Claude debugged it. Claude helped us ship faster than any team our size should be able to.
But Claude didn't make $200K this month.
Infrastructure did.
And that's the part nobody on your timeline is explaining. So let me.
1. What Claude actually did for us.
Let's give credit where it's due.
Claude Code wrote roughly 50,000 lines of our codebase. Order construction logic. WebSocket handlers. API adapters for 12 different prediction market platforms. Error handling. Logging. The boring stuff that takes weeks to write manually.
Claude did it in days.
We used Claude to debug execution edge cases at 3am. We used it to refactor our Jito bundle signing flow. We used it to write the adapter layer that normalizes order formats across Polymarket Kalshi Drift Limitless Myriad and 7 other venues.
Claude is an incredible engineering tool. We wouldn't have shipped in 4 months without it.
But here's what Claude didn't do.
Claude didn't give us private RPC nodes with sub-10ms latency to Solana validators.
Claude didn't give us gRPC Yellowstone streaming at 1-5ms when everyone else is on 200ms WebSocket.
Claude didn't give us Jito atomic bundles that guarantee both legs of a trade fill or neither does.
Claude didn't give us ShredStream access that sees block data 200-500ms before normal propagation.
Claude didn't give us 4 months of testing edge cases across 12 venues at 3am until the system could run 200+ trades per day without a single failed transaction.
Claude wrote the code. We built the infrastructure the code runs on.
The code is 1% of the edge. The infrastructure is the other 99%.
That's the part that made $200K this month. And that's the part no Claude prompt will ever give you.
2. What the system actually does.
Let me show you the trade that generated $200K. Not the code. The mechanic.
Same event lives on two platforms. "Will the Fed cut rates in June?"
Polymarket price: 60% chance. YES costs $0.60. NO costs $0.40.
Kalshi price: 52% chance. YES costs $0.52. NO costs $0.50.
We buy YES on Kalshi for $0.52.
We buy NO on Polymarket for $0.40.
Total spent: $0.92.
If the Fed cuts rates:
YES pays $1.00. We made $0.08.
If the Fed doesn't cut:
NO pays $1.00. We made $0.08.
Both outcomes. Same profit. No prediction needed.
That's not AI being smart. That's two platforms disagreeing on the price of the same event. And we stand in the middle.
This happens on 15-30% of all prediction markets. Every day. Politics. Crypto. Sports. Economics.
The spread is right there. Open Polymarket in one tab. Kalshi in another. Compare prices on the same event. You'll see it within 5 minutes.
Seeing it was never the hard part.
3. Why seeing the spread and capturing the spread are completely different problems.
This is where every Claude bot thread falls apart. And this is where our $200K came from.
The spread is visible to everyone. A 6% gap between Polymarket and Kalshi on a Senate race. An 8% gap on a Fed decision. A 5% gap on a Champions League match.
Anyone can see it.
Almost nobody can capture it. Here's why.
Speed. Cross-platform spreads last seconds to minutes on liquid markets. The average intra-platform window is 2.7 seconds and shrinking. 73% of all arbitrage profits go to systems executing in under 100 milliseconds. A Claude bot takes 1,000-5,000ms just to sign one order on one platform. By the time it's done thinking the opportunity is gone.
Atomicity. You must fill both legs or you're not doing arbitrage. You're making a directional bet. If you buy YES on Kalshi and the NO on Polymarket doesn't fill you're now exposed to the outcome. That's not risk-free. That's gambling. You need atomic execution. Both sides fill or neither does. That requires Jito bundles on Solana. Not a Python script.
Multi-venue connectivity. You need live real-time connections to 12+ platforms simultaneously. Each with its own API. Its own authentication. Its own fee model. Its own order format. All normalized. All monitored 24/7. That's not a weekend project.
Claude can understand all three of these problems perfectly. Claude can explain them better than most engineers.
Claude cannot solve any of them. Because they're not code problems. They're infrastructure problems.
4. What happens in 90 milliseconds.
Let me walk you through a single trade from this month. Real execution.
0ms — Spread detected. Fed Rate Cut market. Polymarket 60% vs Kalshi 52%. Spread: 8%.
15ms — Order pair constructed. YES at $0.52 on Kalshi. NO at $0.40 on Polymarket. Two legs. One atomic bundle.
35ms — Jito bundle signed. All-or-nothing guarantee. Both fill or neither fills. Tip attached for validator priority.
55ms — Bundle submitted via dedicated RPC. Sub-10ms to the validator. Not through a public endpoint. Through our bare-metal node.
75ms — Both sides filled on-chain. Positions confirmed.
90ms — Profit locked. +$0.08 per contract. Guaranteed regardless of outcome.
Total elapsed: 90 milliseconds. Faster than you can blink.
During those 90ms a Claude bot would still be waiting for the API to respond to its first prompt.
We ran this trade hundreds of times a day. Across hundreds of markets. On 12 platforms. For 30 days.
That's where $200K came from.
Not from Claude being smart. From infrastructure being fast.
5. The stack that made $200K.
Solana base layer. ~400ms blocks. Finality approaching 150ms. 65K+ TPS. Sub-$0.01 fees. Both Polymarket (via Jupiter) and Drift are native on Solana now. One chain. One wallet. No bridging.
Dedicated RPC. Bare-metal nodes. Sub-10ms to validators. Stake-weighted QoS. Our transactions get priority during congestion. This alone is the difference between a filled order and a missed opportunity.
gRPC Yellowstone + ShredStream. 1-5ms data updates. Standard WebSocket is 50-200ms. We see orderbook changes 50-100x faster. ShredStream gives us block data 200-500ms before normal Turbine propagation. When your bot is processing old data we're already executing on new data.
Jito Atomic Bundles. Up to 5 transactions. All or nothing. Both legs of every cross-platform trade execute together or don't execute at all. 95% of Solana validators run Jito. Over 3 billion bundles processed per year. This is what makes the trade mathematically risk-free at the execution level.
Cross-Platform Engine. 12+ venues. Normalized adapters. Native WebSocket on every CLOB. 24/7 scanning. When a spread exceeds the threshold both sides fire in ~90ms.
Claude helped write the code for all of this. Claude did not provide any of this.
The code runs anywhere. The infrastructure runs on specific hardware with specific network connections at specific latencies that took months to set up and test.
That's the 99%.
6. What happened in January and why it matters.
In January 2026 Polymarket removed the 500ms taker delay and introduced dynamic fees. On 50-cent contracts fees reach approximately 3.15%.
Every single-platform taker bot died overnight.
The viral Claude bots. The $313 to $438K wallet. The "build your own arb bot in 40 minutes" crowd. All dead. The fee now exceeds the typical intra-platform arbitrage margin. The math stopped working.
We didn't lose a dollar.
We didn't change a single parameter.
Because we don't operate on a single platform. We operate between platforms. Polymarket's fee change affects trades within Polymarket. It doesn't affect the spread between Polymarket and Kalshi. Those are driven by regulatory fragmentation and capital friction. Not by one platform's fee schedule.
The fee change that killed 92.4% of traders didn't touch us.
That's what infrastructure means. You don't depend on any single platform's rules. You operate above them.
7. The lie in every Claude bot thread.
Every viral thread has the same structure.
"Claude made
$X."
"Here's the strategy."
"Here's the prompt."
"Build your own."
"DM me GUIDE."
Here's what they don't tell you.
Claude didn't make
$X. A developer made
$X using code Claude helped write. The code is the visible 1%. The invisible 99% is everything underneath.
The prompt is useless without the infrastructure. You can have the perfect strategy and the perfect code. If your execution takes 3 seconds and the window is 2.7 seconds you lose money. If you can't guarantee both legs fill you lose money. If you're on a public RPC at 500ms latency you lose money.
The strategy is public. The code is open source. The infrastructure is the moat.
That's why 92.4% lose. They have the strategy. They have the code. They don't have the infrastructure.
We spent 4 months building it. So you don't have to.
8. How to use ours.
Everything above. The 12-platform scanning. The Jito bundles. The 90ms execution. The dedicated RPC. The gRPC streaming. The atomic cross-platform engine.
You don't build it. You plug in.
Step 1 — Go to polyarb.io Step 2 — Create account. Email and password.
Step 3 — In Jupiter Prediction settings replace the default RPC with: https://polyarb-rpc.solana.com Step 4 — Set your parameters. Choose platforms to scan. Set spread threshold. Choose automated or alert mode.
Step 7 — Activate. It runs 24/7. You watch the P&L. Or don't. It doesn't need you.
Cost: $0 upfront. Profit-sharing only. We make money when you make money.
Setup: under 5 minutes.
Then it runs whether you're awake or not.
9. So does our Claude bot make $200K?
Yes. The system made $200K this month.
But calling it a "Claude bot" is like calling a Ferrari "a wrench project" because someone used a wrench to build it.
Claude is the wrench. An incredible wrench. The best wrench most of us have ever used.
But the wrench didn't win the race. The car did.
The car is the infrastructure. The private nodes. The atomic bundles. The 90ms execution across 12 platforms. The 4 months of engineering. The 50,000 lines of code that Claude helped write but could never run.
Every week a new thread goes viral telling you to build a Claude bot on Polymarket.
We're telling you something different.
Stop building the wrench. Drive the car.