How $1 per minute is calculated
An agentic work cycle consumes significantly more tokens than a simple chatbot query. The agent reads context, plans, calls tools, processes results. Every step adds tokens.
Using Claude Sonnet 3.7 (the model I used in the first part of the experiment), the average cost of an active agentic cycle worked out to roughly $0.80–1.20 per minute of actual work. Average: approximately $1/min.
But the agent doesn't work continuously — it's closer to the 15-minutes-per-hour ratio mentioned above. So an hour of the total process actually costs around $15, not $60.
Where $9,600 per month comes from
A simple calculation: 8 hours per day, 5 days per week, agent active for 15 minutes per hour.
- Active minutes per day: 8 × 15 = 120 minutes
- Cost per day: 120 × $1 = $120
- Working days per month: ~20
- Monthly cost: $120 × 20 = $2,400
At full agent utilisation without optimisation you can reach $2,400–4,000 per month. The $9,600 figure appears with a less optimised approach or when using a more powerful model — and I've seen it at teams that aren't watching the orchestration closely.
Comparison with a senior developer
A senior developer in Western Europe costs roughly €3,500–5,000 gross per month. Add employer contributions, benefits, onboarding time, holidays and sick days.
At first glance the AI agent looks comparably expensive. But the comparison is misleading in both directions:
- The agent runs 24/7 — given work, it won't stop
- No context switching — every cycle starts clean
- No code review — that stays with the human
- No ideas — it responds to assignments, doesn't create strategy
Realistic conclusion: agentic AI combined with one senior developer replaces a team of two to three mid-level developers on repetitive tasks. Strategic decisions and architecture stay with the human.
Local models as an escape route
On day three I seriously considered local models for the first time. API pricing is manageable for a pilot project, but scaling it to a full team is a different number entirely.
Local models — Qwen, DeepSeek, Mistral — offer zero inference cost after paying for hardware or a cloud instance. Their quality has closed the gap with proprietary models significantly over the past year.
More on local models in practice comes in day 5 — but the first seed of doubt about cloud API dependency was planted right here.
Want to see how business processes can be automated? Book a consultation — we start where vibe-coding ends.
In the next episode
Day 4: What happens when Anthropic bans you. Not theory — it actually happened.