People always ask about the cost
Every time I mention that I run 26 AI agents, the first question is "how much does that cost?" Fair question. Most people assume the answer is astronomical. It's not. But the breakdown might surprise you, because the expensive parts aren't where you'd expect.
Let me walk through the real numbers.
The hardware: the big upfront investment
My Mac Studio is the foundation of the whole setup. M-series chip, 98GB of unified memory. It cost somewhere north of 4,000 USD when I bought it. That's a real amount of money. I'm not going to pretend otherwise.
But here's what that 98GB of unified memory buys me: the ability to run Llama 3.3 70B, Qwen 32B, and GLM Flash locally via Ollama. These are real, capable language models running on my own hardware with zero per-token cost. Once you own the machine, every inference is free.
The Mac Studio runs 24/7. It hosts the agents, runs the local models, serves development environments for all my projects, and handles the Ollama inference. The electricity cost in Oslo is modest. The machine is quiet. It sits on my desk and works.
If I were starting from scratch today, the Mac Studio would still be my first purchase. It's the single investment that makes the rest of the cost structure manageable.
What runs locally (and it's a lot)
About 70% of all the work my agents do runs on local models. Free.
Synapse, my operations agent, runs on Llama 3.1 8B. It handles cron jobs, cleanup tasks, knowledge synchronization between agents. None of this needs the best reasoning model in the world. It needs reliability and availability, which a local model provides.
The local models handle text classification, summary generation, routine analysis, document parsing, and all the maintenance work that keeps the system running. When Hermes needs to classify an email, it doesn't need Opus. It needs a fast model that can say "this is a receipt" or "this is a newsletter" quickly and cheaply. Local models do that well.
I run the local models through Ollama, which gives me a simple API that works the same way as the cloud model APIs. The agents don't know or care whether they're talking to a local model or a cloud model. The abstraction is clean.
The 70% number isn't a guess. I've tracked it. Most of what agents do day-to-day is routine: reading files, checking statuses, syncing knowledge, processing structured data. This work doesn't need Claude or GPT. It needs a model that's available, fast, and free.
Cloud model costs: the 30% that matters
The other 30% of the workload goes to cloud models, and this is where actual money gets spent.
Anthropic is the biggest line item. Claude Sonnet 4.6 is the default model for most of my agents. Nyx, Hermes, Sentinel, the product agents, the personal agents. Sonnet is capable enough for strategic thinking, writing, and analysis, and it's significantly cheaper per token than Opus.
Claude Opus 4.6 is reserved for one agent: Forge. Code generation and architecture decisions need the strongest reasoning available. When Forge is writing a complex React component or designing a database schema, the difference between Sonnet and Opus is noticeable. Opus catches edge cases that Sonnet misses. It produces better-structured code. For engineering work, it's worth the premium.
Gemini 2.5 Flash handles batch processing and research. Scout and Atlas both run on it. Research work involves a lot of reading and summarizing, which Flash does well at a fraction of the cost of Anthropic models. When I need to process a large batch of data, like indexing emails or running analysis across multiple sources, it goes to Gemini.
GPT-4o-mini picks up the marketing agents. CMO, SEO, Email Marketing, Ads, Partnerships. These agents do structured work: drafting copy variations, analyzing keywords, formatting campaign data. GPT-4o-mini is cheap and reliable for this kind of task.
My monthly API spend varies depending on how much engineering work Forge does, because Opus tokens add up faster than anything else. A month where I'm actively building a product costs more than a month of steady-state operations.
The cost I didn't expect: service APIs
Here's what surprised me. The AI model APIs aren't the most annoying cost center. The service APIs are.
Gmail API calls for Hermes to process 83,000+ emails. Supabase for seven different project databases. Fiken API for accounting. Stripe for payment processing in EventRipple. GitHub API for PR management. Each of these has its own pricing structure, rate limits, and authentication complexity.
Supabase offers a generous free tier per project, but I have seven projects. Some of them push past the free tier thresholds. The database usage for Budget, EventRipple, and PortLink adds up when you're storing transactions, events, and maritime data.
The real cost of service APIs isn't just the dollar amount. It's the operational overhead. Managing OAuth tokens, handling rate limits, dealing with API changes, monitoring for failures. Every external integration is a surface area for something to break. When Gmail changes something in their API response format, Hermes needs updating. When Fiken deprecates an endpoint, the accounting pipeline needs fixing.
If I had to estimate, I'd say the time I spend maintaining service integrations exceeds the time I spend on AI model configuration by a factor of two or three.
The rate limit lesson
This cost me time and sanity more than money, but it belongs in a cost breakdown because it affects how I allocate spending.
About four weeks in, I had multiple agents making Anthropic API calls simultaneously. Five agents, all hitting Claude's API at the same time. The API started returning 529 errors. Overloaded. Everything stopped. Agents were stuck mid-task with partially written files.
The fix was a hard rule that's now broadcast to every agent: maximum two Anthropic API calls in parallel. Ever. No exceptions. Batch work goes to Gemini Flash. If you hit a 529, switch to a non-Anthropic model immediately. Don't retry in a tight loop.
This rule shaped my cost structure. Heavy processing jobs that could theoretically run on Claude get routed to Gemini instead, not because Gemini is better for the task, but because I can't afford to saturate my Anthropic quota during business hours when Nyx and Forge need it.
I also schedule intensive Anthropic jobs between 2am and 6am Oslo time to avoid peak hours. The Folio sync runs at 2am. Knowledge synchronization runs at 3am. By the time I wake up, the heavy work is done and the rate limits are fresh.
What I'd optimize
If I were rebuilding the cost structure from scratch, I'd change a few things.
I'd be more aggressive about routing work to local models from the start. In the early weeks, I defaulted too many agents to cloud models because I didn't trust the local ones. After testing, Llama 3.3 70B handles most non-creative, non-reasoning tasks perfectly well. The quality difference for routine work is negligible. The cost difference is 100%.
I'd pick one cloud provider for batch work and stick with it. I have tasks scattered across Gemini, GPT-4o-mini, and occasionally Anthropic for batch processing. The fragmentation means I'm managing three different API integrations, three billing dashboards, three sets of rate limits. If I consolidated batch work entirely on Gemini Flash, the operational simplicity would be worth any marginal cost difference.
I'd invest more upfront in caching and deduplication. Some of my agents ask similar questions repeatedly because their memory files don't capture every interaction. Better memory management means fewer API calls, which means lower costs. The cheapest API call is the one you don't make.
A realistic monthly estimate
Let me put rough numbers on this. These are my actual ranges, not theoretical calculations.
The Mac Studio was the one-time capital cost. Amortized over two years, call it around 170 USD per month. Your mileage varies depending on what hardware you already own.
Anthropic API (Sonnet + Opus combined) is the biggest recurring cost. The range depends heavily on how much engineering work Forge does. Light month, it's modest. Heavy development month with Forge building features on Opus, it goes up considerably.
Gemini and GPT costs are lower. Gemini Flash for batch and research work is cheap by design. GPT-4o-mini for marketing agents is pennies per task.
Supabase across seven projects with some on paid tiers adds up, but it's predictable.
Other services: Netlify for deployments, GitHub, domain registrations, Slack (which is free for my use case). These are individually small but collectively noticeable.
Local model costs are electricity only, which in Norway is a rounding error on my power bill.
Total monthly: somewhere in the range of a nice dinner out to a weekend trip, depending on how hard I'm pushing the cloud models. It's not nothing. It's also not what most people expect when they hear "26 AI agents."
Is it worth it compared to humans?
This is the question that actually matters. Not "is it cheap?" but "is it worth what I pay?"
Let me frame it differently. My agents handle: email management and CRM for 83K emails. Code generation and review for five active products. Marketing strategy, content creation, and SEO for three brands. Personal finance management. Research and competitive analysis. Home automation. Fitness tracking. Accounting pipeline.
If I hired humans for this work, I'd need at minimum: a developer, a marketing person, a personal assistant, and a bookkeeper. In Oslo, that's easily 30,000-40,000 USD per month in salaries alone. Plus management overhead, office space, equipment, benefits.
My setup costs a fraction of that. Yes, the agents don't replace senior humans one-for-one. I still make all the decisions, do all the strategy, handle all the client relationships. But they handle the execution work that would otherwise require me to either do it myself (slow) or hire people to do it (expensive).
The comparison isn't perfect. A human developer has judgment that Forge doesn't. A human marketing person reads social cues that the CMO agent can't. But for a solo operator running multiple projects, the agent setup gives me capacity I couldn't otherwise afford.
The real cost is your time
I haven't mentioned the most significant cost, and it's not financial. It's the time investment.
Setting up 26 agents took weeks. Writing memory files, configuring tools, establishing communication protocols, debugging integration issues, iterating on standing orders. The first month was as much about building the system as it was about using it.
Ongoing maintenance is real. Memory files need pruning. Agent configurations need updating when projects change. New integrations need building. When something breaks, I'm the one fixing it. There's no support team to call.
I estimate I spend 5-10 hours per week on system maintenance and improvement. That's time I'm not spending on actual product work. For a solo founder, that's a significant allocation.
The payoff is that the other 40+ hours of my work week are amplified by the system. But the investment period is real, and if you're evaluating whether to build something like this, you should account for it honestly.
The bottom line
Running 26 AI agents costs less than you'd think for the infrastructure and more than you'd think for the maintenance time. The hardware investment pays for itself through free local inference. The cloud model costs are manageable if you're strategic about routing work to the right model tier. The service API costs and integration overhead are the real budget items that don't show up in the marketing materials for any AI platform.
For a solo founder running multiple projects, the math works. Not because it's cheap in absolute terms, but because the alternative, doing everything myself or hiring a team, is either slower or dramatically more expensive.
The 26-agent setup isn't for everyone. If you're running a single project with a clear scope, you probably don't need this. If you're juggling multiple products, multiple clients, and multiple business functions simultaneously, the multiplied capacity is hard to get any other way.
The numbers are real. The tradeoffs are real. And the 70% of work handled by free local models is the number that makes the whole thing viable.