I Audited My SaaS Subscriptions. AI Tools Are Now 60% of My Stack Cost.
I Audited My SaaS Subscriptions. AI Tools Are Now 60% of My Stack Cost.
Once a quarter I sit down with a spreadsheet and list every monthly subscription that hits a card I own. The first time I did it, in late 2023, the exercise took twenty minutes and the AI category was a single line for ChatGPT Plus. The audit I did this weekend took two hours, and AI tooling — Claude, Cursor, Codex, three API keys, and assorted small things — accounts for roughly 60% of the total monthly burn.
That number is honest. It is also not what I expected. I'd been telling myself I was being disciplined about AI spend. The audit said otherwise. So I cut three subscriptions, kept some I would have cut a year ago, and I'm now thinking differently about how I budget for tools. This is the post-mortem.
The breakdown
I'll spare you the exact dollar figures because (a) some are work expenses I don't want to publish on the open internet, and (b) the absolute numbers matter less than the proportions. The shape:
AI tooling: about 60% of monthly burn. The breakdown inside that is roughly 40% paid plans (Claude Max, Cursor, ChatGPT, GitHub Copilot for one specific repo), 35% API spend (Claude API for tool use in side projects, OpenAI for embeddings and Whisper, a small amount of Gemini), and 25% the long tail of "AI-powered" SaaS that wraps a model in a UI (transcription, image generation, a writing tool, a research tool).
Hosting and infra: about 18%. Cloudflare, Vercel, Supabase, Backblaze, a couple of domains. This category has been remarkably stable for years and I don't expect it to change.
Productivity and ops: about 12%. Notion, 1Password, an email tool, a calendar tool, a couple of small subscriptions for accounting and invoicing.
Design and creative: about 6%. Figma, a stock asset license, an icon set.
Everything else: about 4%. Various $5-15/month tools that exist for one specific reason and earn their keep.
The shape itself is the data point. Two years ago, hosting was the biggest line. Now AI is hosting plus design plus most of productivity, combined.
The three subscriptions I cancelled this month
I cancelled three things during the audit. The decision math on each was different.
The first was a research-assistant SaaS that wraps Claude with a web-search tool. I'd been paying for it because it was convenient — a clean UI, a saved-searches feature, and it remembered context across sessions. The audit forced the question: convenient compared to what? I do the same task with the Claude desktop app's web search and a manual notes file, and the convenience premium I was paying turned out to be about $30/month. Gone.
The second was a transcription service I used to use weekly that had drifted to monthly use, then to "I haven't opened this in a quarter." The honest answer was that Whisper via the OpenAI API does the same job for cents per hour, and I'd built the script to do it months ago and then forgot to cancel the SaaS. Classic stale subscription.
The third was harder. A code-search tool I paid for because it was genuinely useful — better cross-repo search, AI summaries — but that I now use less since Claude Code can read and understand whole repos directly. The tool didn't get worse. The base level capability moved up around it. The cancellation here wasn't "this isn't worth $25/month" — it was "the value of $25/month of code-search has dropped to $5/month of code-search because the surrounding stack improved."
What I'm overpaying on, on purpose
Two line items in the audit looked overpriced and I left them alone.
Claude Max. The pricing is real money, and I could probably get by on the cheaper plan if I were strict about my usage. But the cost of being wrong about the limits, mid-flow, on a deadline, is higher than the price difference. I'm paying for the ceiling, not the average usage, and that's a deliberate choice.
Hosting redundancy. I keep this blog deployable to two different platforms with about an hour of setup. That redundancy is paid for in extra config and a little bit of duplicate cost. After the Vercel-Context.ai breach earlier this year, paying a small premium to not be locked into one provider feels appropriate. I'd rather have the optionality than the saved $10/month.
Both of these violate the "if you don't use 80% of the value, cut it" heuristic. I left them anyway. The audit's job is not to enforce the rule — it's to surface the violations and let me make a real decision on each one.
Where I'm underpaying and regretting it
Two things were underspending I noticed on the way through.
Backups. I run nightly backups on the things that matter, but my offsite tier is leaner than it should be for the value of the data. I'm planning to upgrade Backblaze tier this month. The cost of being wrong about backups — losing data — is asymmetric. I should be paying more, not less, for the safety margin.
Accounting software. I've been using a $9/month tool that technically does what I need but that I dread opening. The friction of using it means I do my books less often, which means I make worse business decisions because I have less timely data. A $40/month tool I'd actually open is a better deal even though the price is higher. Trying one this quarter.
The pattern in both: I was optimizing for monthly cost and not factoring in the cost of being wrong, or the cost of friction. Both of those are real costs and they don't show up on the credit card statement.
The rule I use now
The rule that emerged from this audit is simple to state and useful in practice. For every $20/month tool, ask yourself the question: if this disappeared overnight, would I rebuild it, or would I just stop doing the thing it does?
If the answer is "I'd rebuild it" — even crudely — the tool is doing real work for me and the price is fine. That's an objective measure of value: I am willing to spend my time replacing what the tool does.
If the answer is "I'd just stop doing the thing" — the tool isn't earning its slot. I was doing the activity because the tool was there, not because the activity was actually important. Cancel.
This is a sharper test than the standard "do you use it weekly" heuristic, because it forces you to confront whether the underlying activity matters at all. A lot of the SaaS I've cancelled over the years was valuable software for an activity I didn't actually need to be doing.
What I'd actually do today
If you haven't done a stack audit in the last six months, do one this weekend. Two hours, a spreadsheet, every charge on every card. Most solo devs I know who do this find at least one stale subscription and one underpriced thing they're missing.
Apply the "would I rebuild it" test to every line item. Be honest about the answer.
Watch your AI spend specifically. The category has grown faster than any other line item in my budget for two years running, and it's the line that needs the most attention precisely because the per-tool cost is small enough that you stop noticing. The $20 here and $30 there compounds.
The audit is not about being cheap. It's about making sure the money you spend on tools is actually buying you output, not buying you the feeling of being well-tooled. Those two things look identical until you stop and count.