Most "AI is expensive" discourse is vague. The pieces that quote real numbers usually quote enterprise tier list prices for tools the writer does not personally run. The pieces that talk about personal use rarely quote any numbers at all.
This is the ledger for a working personal AI stack that ships a five-surface daily content pipeline, runs four cron jobs, holds a SQLite memory database, and supports about thirty published pieces a month. Real line items, monthly recurring, in the rough order of largest to smallest.
The Numbers
Claude Max subscription: $200/month. This covers Claude Code, the writing model for every draft, the editor on every revision, the OAuth identity for tool integrations. No separate API key needed for Pro/Max users on most workflows. The two-hundred-a-month tier gives me the rate limits I need for daily use; the lower Pro tier at $20 ran out of capacity inside the first heavy week.
Anthropic API spend, outside Max: roughly $30–60/month. Used for the cron-driven trend-scan, digest, and draft pipeline that runs without an interactive Claude Code session — Sonnet for T1/T2 drafts, Opus for T3, occasional Haiku for the lint-eval cycle. Spend tracks topic volume. Slow weeks land near $30; busy weeks with deep research on every piece push toward $60.
Late.dev paid plan: about $30/month. Hit the 20-post free cap on April 26. The upgrade was immediate because the alternative was bifurcating publishing across two manual flows for LinkedIn/X/IG. Current usage running 60+ posts a month across the three social surfaces means about $0.50 per post in distribution cost.
Ghost Pro hosting: $25/month. The standalone option for self-hosting at lower cost on a VPS exists, but Ghost Pro covers backups, CDN, email delivery, and admin auth for less than the time cost of running it myself. The five-surface piece always starts at Ghost as the canonical URL.
Firecrawl: $20/month. The base plan covers the trend-scan crawls and the per-piece research lookups (≥3 sources per T3 draft). Usage tracks topic generation rate, not piece count. Slow research months stay under the cap; weeks with five separate contrarian sources per piece can push over.
Dev.to, Cloudflare, GitHub, Cursor (for occasional sidecar work): $0–20/month. Dev.to is free for publishers. Cloudflare on the free tier handles DNS and Access for the dashboards. GitHub Free for personal repos. Cursor I use for one specific kind of work outside Claude Code; the free tier has been enough this month.
NAS storage and home server: roughly $15/month amortized. A Synology DS920+ bought outright in 2023, running the canonical content-engine path mount and a few Plex services. Cost is electricity plus a notional five-year amortization of the hardware. Not strictly AI spend; without the NAS the content DB would live on a $5/month VPS instead.
Domain registrations (mpt.solutions, mpt.codes, a couple others): about $5/month amortized. Annual renewals divided by twelve.
The Total
Run the line items: $200 + $45 + $30 + $25 + $20 + $10 + $15 + $5 ≈ $350/month, with seasonal variance pulling it to roughly $320–420 depending on usage.
That is the gross. The net story is more interesting.
What This Replaces
Look at the equivalent enterprise plan that would deliver the same operator experience.
Microsoft 365 Copilot Business is $21/user/month, but that covers AI in Word and Excel and Outlook — none of which is the agentic coding loop. Add ChatGPT Team at $30/user/month for the writing side. Add Cursor Business at $40/user/month for the coding side. Add an enterprise scheduling tool like Buffer for the social fanout at $100/month. Add a CMS subscription at Ghost-or-equivalent rates. Add an enterprise search API like Tavily or SerpAPI for the research crawls at $100/month.
That assembled stack runs roughly $300/user/month for the licenses alone, and produces an experience that does not include any of: persistent cross-session memory, lint enforcement against my voice guide, queue-driven scheduled publish, custom hooks against my git workflow, or the SQLite schema that makes the dedupe and the cross-platform reconciliation work. Those parts would still need to be built on top.
The personal stack costs less than the enterprise license stack and does more, because the wrapper is mine and the wrapper is where the payoff lives. This is the compose-the-stack argument from May 21 in dollar form. The vendor positioning charges you for the model and the surface. The personal stack pays the vendor for the model only, and you build the surface.
Where the Spend Lands Wrong
Honesty about waste matters more than the gross number.
The Anthropic API spend on the cron-driven pipeline is the line item with the worst yield. The AI-generated drafts get used about three days out of ten — the other seven, I write the piece by hand and run only the publisher path. The cron pipeline costs about $30/month in API calls to produce drafts that mostly get discarded in favor of human writing. I keep it running because the digest output is useful even when the drafts are not, but the marginal API call against an Opus draft I will not ship is the easiest line item to defend cutting.
The Firecrawl cost runs higher than it needs to because the trend-scan queries are overlapping — TechMeme returns about 40% the same stories as the Hacker News pull, and the Reddit subreddit list is wider than it needs to be. A focused trend-scan would hit the same signal at half the credit cost.
Late.dev is the line item with the worst risk profile. Single-vendor dependency on a fast-moving startup for three of the five surfaces. The May 2 IG 409 false-negative incident was an example of where the vendor's behavior diverges from documented contract, and the cost of switching is rewriting the publisher integration in publisher.ts. Not painful, but real.
Where the Spend Lands Right
The Claude Max subscription is the line item with the largest payoff by an order of magnitude. The work done in interactive Claude Code sessions — the actual writing, the actual drafting of these posts, the actual debugging of the publisher integration, the actual building of the personal stack itself — is what produces value. Cutting $200 there would tank the entire output. Doubling it would not change the output much, because the bottleneck is what I think and write, not the agent's rate limit.
Ghost Pro is a small line item that has zero failure modes. Self-hosting would save $25/month at the cost of recurring incidents that compound over a year. The premium for not having to think about CMS uptime is the right premium.
The lint-and-publisher infrastructure is free in operating cost and produces all of the consistency value. The compounding piece of this stack is not the line items that cost money. It is the parts I wrote myself that do not.
The Comparison That Matters
A senior engineer at a company that has not yet rolled out an enterprise AI plan can run a personal stack equivalent to mine for about $350/month, build it in a weekend, and ship faster than the team will officially be allowed to ship for the next eighteen months. The math against the eventual enterprise plan that lands in 2027 will look like a rounding error.
A senior engineer at a company that has rolled out an enterprise AI plan can run the same stack for the same $350/month, with the same independence, regardless of whether the official plan is good. The official plan being good is not a precondition for the personal stack working. The official plan being bad is not a reason to wait either.
The cost is real. The payoff is realer. Three hundred and fifty dollars a month is what it costs me to ship at this rate. The exact mix will be different for everyone. The order of magnitude will not be.
