7 months after my "$0/month AI agents" post: what it costs now
Table of Contents
In March I posted on Hacker News that four AI agents ran my one-person company on Gemini's free tier for $0 a month. It got 16 points and a lot of fair hits. One commenter counted 10 replies from me in six minutes and said he didn't believe I wrote them. Another asked for my MRR. I never gave a number.
This is the follow-up, with that number in it. The $0 setup is gone. The main bill is now two Claude subscriptions, about $400 a month, and Google startup credits cover my cloud costs.
Disclosure first this time. English is not my language. Claude drafted this post from my logs and cost sheets, and I edited it. Every number below comes from a log or a bill, except my hours, which are a guess and labeled as one.
Why the free tier didn't last
In March my agents had 1,500 free requests a day on Gemini 2.5 Flash and used about 105. One of my video pipelines still runs on a free key today, and that key gets 20 a day. On Oct 5 an evening of testing plus retries used them up. The next morning's run made 1 video and failed 5, all with 429 errors.
In July a batch job that re-embedded a document library cost NT$1,001 (about $32) in a single day on a billed key. The work was legitimate. Nobody priced it before it ran. Since then any batch job estimated over NT$200 waits for my OK.
The March agents mostly stopped. Their public auto-posting broke on July 10 and I didn't fix it. On Aug 4 I shut down the gateway and cleared all 10 scheduled jobs. One agent had already been switched off earlier because it was posting junk.
What runs it now
The work moved to Claude Code. I pay for two Max 20x accounts. A small script watches for rate-limit messages and swaps credentials when one account hits its weekly cap. The first account hits that cap every week. In late September, ccusage put its usage at $1.9k to $3.9k per week at API list prices. Two accounts are still not enough: on Oct 11 both ran out of Opus and everything sat idle for about 20 hours. I know this is a subsidy I can't plan around.
One long-running session acts as the coordinator. At start it reads a few operating docs: a facts table where every line has a last-verified date, notes on when to stop and ask me, and eight hard rules. Half of those rules have the incident that caused them written right next to them. The coordinator spawns headless claude -p workers for real tasks. A separate critic agent attacks anything that will be seen outside.
The home PC has an RTX 3080. It runs local TTS (BreezyVoice, a Taiwanese-accent voice model), video rendering, and image generation, at zero marginal cost. Recurring jobs moved off Claude Code's in-session cron to Windows Task Scheduler, because session crons quietly expire after 7 days. On Oct 1 a patrol job had been dark for 15 hours before anyone noticed.
The leftovers of the old fleet ran on OpenRouter. Lifetime spend there is $59.73, and the balance is now close to zero, so a couple of those jobs have been failing since Oct 1. Telegram got demoted too. Since Sept 3 it only carries alerts, and I read the coordinator's output on my phone through Remote Control. I also built an ESP32 board to wake the PC while I travel. Its offline alert passed a test in September. Wake-on-LAN was never tested, and when I checked on Oct 12 the board wasn't even online.
Google Workspace costs $26.40 a month after I cut a seat, and Vercel about $22. Google Cloud runs about NT$4,300 a month, paid by startup credits that expire in July 2027.
What I cut
I froze the Pin runtime, a consumer skill bot for LINE, Telegram and WhatsApp, on Sept 3. Only the self-serve site builder survived. Months earlier I had written down a test for myself: if a revenue line needs me present to work, it's dead. I turned off cold email and Bluesky auto-posting. Three side repos (a trading tool, a static site, a 360-video tool) had zero commits in September and are being archived.
I tried local LLMs for writing. qwen3.5 9B, run on CPU, wrote 7 YouTube descriptions with 0 word-choice errors in 57 sentences. The facts were another story: 4 of the 7 had to be rejected, and one copied an internal reference code into a public description.
I tried a cheaper model for drafting. In a blind test on 7 finance scripts, Sonnet cost 0.66x as much as Opus and its average score was slightly higher. It also made 4 factual errors to Opus's 1. The comparison wasn't perfectly fair, since the Opus scripts had already been through one revision. On insurance and tax content, 4 factual errors decided it. That experiment cost about $47 at API list prices. Before it, two rounds with a fancier HTML video framework cost $35, and when I watched the results I couldn't see a difference.
Where the money and time go
Nine YouTube channels exist and a tenth is planned. Six are "how it works" channels: machines cut open, kitchen chemistry, weapons mechanics, cameras, instruments, art. They are meant to send traffic to the products. I don't expect ad revenue. Each video is written, rendered and reviewed by agents, and a weekly loop around that (metrics, a playbook per channel, topic picks, production, mechanical checks, a critic, scheduling) goes live on Oct 19. So far it has had four trial runs. Two videos were approved and two were held by the critic.
Only one of the four passed review on the first try. It took about 250k tokens, roughly $3 at API list prices. The other three took 390k to 510k. I built a cheaper path for wording-only fixes but haven't measured it yet. Two pipelines share one GPU, so a lock file makes them take turns. On the first real run a re-render waited 9 minutes in line instead of being skipped. By design I step in when a new channel starts, when the weekly report asks me a question, or when the critic blocks a video.
Results so far are close to nothing. The first video on the machines channel went public on Oct 13 and had 10 views an hour later.
MindThread is my Threads scheduling SaaS on Meta's official API. MRR is NT$24,890 (about $785) from 11 paying accounts. September brought in NT$24,580 gross from 12 orders. In the last 90 days 315 people signed up and 13 paid, 4.1%. The biggest leak is the 100 people who created a profile and never connected a Threads account.
Then I broke billing. On Oct 2 I moved monthly plans to recurring checkout, and almost no regular customer could get through it. On Oct 8 I closed recurring checkout and went back to one-time payment links.
Separately, one company pays a monthly fee to run our products as its AI department.
I don't log my hours. On Oct 12 my messages to the system landed in about a dozen different clock hours, mostly one-line decisions. My guess is 2 to 4 hours of real attention a day, in pieces. This is not my full-time job.
Some things stay with me. Anything that spends money or changes a price waits for my OK. Customer replies are drafted, checked by the critic, and then I decide and send. One reply this week came back from the critic with corrections before I sent it. Agents can't use a real card, so I test payment paths with mine and refund. My insurance content has to cite the source text, and those scripts stay on the expensive model.
I still don't have a clean rule for cutting things. Besides the Pin test, I learned one the hard way: before you kill something for zero revenue, check that people can actually pay. My handbook sold zero copies for a month before anyone noticed the checkout redirect was stuck. My growth product went 7 weeks with zero payments before I found the same kind of break, so I still don't know whether zero meant no demand.
Three problems I haven't solved
Renewals are still manual. With recurring checkout closed, only 1 of the 11 paying MindThread accounts renews automatically. The other 10 have to come back and pay again each period.
My worst failures this month were jobs that stopped running and reported nothing. Two incidents on Oct 2 were found by customers before my monitoring saw them.
I don't know whether AI-made video is worth anything. YouTube's inauthentic content policy, in force since July 2025, targets mass-produced templated video, and the six explainer channels are templated. They have no voiceover, so the synthetic-media flag I set on voiced uploads doesn't apply to them. The research in each script is ours and the sources are listed. Whether that is enough, and whether anyone wants to watch, I can't tell from 10 views.
I'll answer questions on Hacker News, slower than last time.
The March post: Hacker News discussion. This week on the video pipeline: Letting agents produce our YouTube Shorts, Three experiments to cut our AI video pipeline's cost, Several Claude sessions, one GPU.