If anything saves me money, it sounds like music to my ears. I'm an average guy in my early 30s. I can't help it.
I've spent an absolute fortune on AI tools. Claude, Codex, Deepseek, GLM, Kimi, Opencode, routing models on Hermes Agent, Pi Agent, hacking Codex to handle third-party models... you name it, I've tried it.
I've tested virtually every AI tool on the planet. It pains me to admit that I've poured lakhs into this space. But to me, it was worth it.
If you bring up Gemini 2.6 Flash, I can tell you its exact strengths and weaknesses off the top of my head. Not because I read a benchmark paper, but because I've actually hammered it in production. A million times over.
At this stage, whether it's a flagship frontier model or a budget LLM, it knows me as well as it knows its own training data.
The Tradeoff
The tradeoff for all this knowledge is pure cash. You have to burn money to truly master AI. I still remember Sam Altman's tweet inviting everyone to try ChatGPT. That was just the beginning. Back then, AI felt practically free. Then Claude happened, prices skyrocketed, and every provider followed suit.
Recently, I bought the Codex Pro plan for ₹10,000. It hit the weekly usage limit in just one day.
Do you realize what that means? That ₹10,000 is supposed to last 4 weeks, meaning a weekly budget of ₹2,500 was burned to the ground in 24 hours.
The Catch
If you know this space, you're probably thinking: "This guy is just illiterate when it comes to token optimization."
Nah, nah, nah. Give me your setup. I'll optimize every single token you own.
My optimization stack is airtight:
- Compressed System Prompts: Stripped down to essential logic.
- Lean Workspaces: Unnecessary skills and MCPs are ruthlessly weeded out.
- Focused Directives: Root-level and project-level claude.md and agent.md files are razor-sharp.
- De-bloated Files: Every skill file is trimmed down to pure substance.
I even set up manual intelligent routing that directs every prompt to its most cost-effective model. If I ask it to calculate 55 × 23, it won't disturb frontier heavyweights like GPT-5.6. Instead, it wakes up a budget workhorse like GPT-5.4 mini.
I could write a book on token optimization. Yet AI is still painfully expensive for someone like me who knows it inside out. Imagine what it's doing to beginners. They're losing lakhs in learning costs alone. I truly feel for them.
The Unexpected Savior
Read this and you'll assume Cursor paid me to write it. (Honestly, I wish Cursor would see this and toss me some free credits or a usage reset.)
It's an incredible tool. I honestly don't know why I slept on it for so long, but its standout feature is something I didn't expect to fall in love with: the usage limits.
They are ridiculously generous. Cursor splits usage into two pools, one of which covers their own model lineup including Cursor 2.5 and Grok 4.5.
Side note: Grok 4.5 is fantastic. Fast, sharp, and highly intelligent. I tried it on the Grok Supergrok plan, and now that Cursor added it to their native model library, they doubled the model pool usage limit. I assumed it was a temporary promo, but they've confirmed the increase is permanent.
I tested it today: I canceled my Codex Pro plan, requested a full refund for the ₹10,000, and switched to Cursor Pro for roughly ₹2,000.
After running it on Auto mode for a full day of heavy work, it only consumed 3% of my monthly quota. That is remarkably fair for the volume of code I churn out.
The Reality Check
I should have made the switch to Cursor much earlier. For a solo founder, it would have saved heaps of time and a small fortune.
As I write this, though, a realistic thought hits me: everyone in AI is out to monetize. Cursor is no exception. Sooner or later, they'll realize how generous they're being and dial back the limits. Until that day comes, I'm going to milk it for all it's worth.
Over to You
How is your AI journey going? Have you started diving in, or are you still holding off?
Eventually, you will start or the market will force your hand.
Shoot me an email at praveensimon@10thlayer.com and tell me your story.
(If someone forwarded this to you, make sure to subscribe so you don't miss the next letter.)
Ciao, Praveen Simon Founder, 10th Layer