My $5,000/month X API bill, and the architecture I should have built first

I spent three months last year trying to build a Twitter sentiment tracker for a client. It was supposed to be a small side project — pull mentions of their brand, sentiment-tag them, ship a weekly report. Maybe a $5K total engagement.

I burned through about $11K of my own time and roughly $800 in API credits before I shipped anything that worked. This is what I learned, and what I wish someone had written before I started.

Phase 1: "I'll just use the free API"

There is no free API for reading tweets in 2026. There hasn't been since February 2023. Most older tutorials still claim there is — they just don't get updated. My first day was spent learning this.

Phase 2: "I'll just snscrape it"

Snscrape's GitHub repo says deprecated at the top. The forks all break on X's current guest-token rotation. I patched one for two days and got 6 hours of usable data before X changed something.

Phase 3: "Fine, I'll buy the API"

I subscribed to Basic at $200/month. The 10,000 read limit was gone in about 6 days of light tracking. Upgraded to Pro at $5,000/month. Suddenly the project was no longer profitable for the client at any reasonable margin.

Phase 4: "I'll build a stealth Playwright pipeline"

Bought residential proxies. Created burner X accounts. Wrote a stealth Playwright scraper with realistic delays and fingerprint patches. Worked beautifully for about 36 hours. Then most accounts got suspended.

What I did not understand at the time: X does not primarily detect on browser fingerprint. It scores account behavior history. A fresh account doing 200 profile views in an hour is suspicious no matter how stealthy your browser is.

Phase 5: "What if I just used my own browser, logged in as me?"

That was the realization. My real X account has history, cookies, normal behavior, and a real browser session. If automation drives that browser to scroll the queries I need, X has much less to detect because there is nothing meaningfully different from a normal browser session.

The lesson is broader than Twitter: most scraping tools are in the business of faking humanity at scale. Faking IP origins, browser fingerprints, and session histories. The architecture that survives is the one where there is nothing to fake.

Full pricing breakdown and failure modes:
https://www.browseract.com/blog/twitter-scraping-2026

Curious if anyone else hit the same wall in 2026 — what did you end up with?