My personal agent is open source, but it is not remotely free.
Between what it consumes directly, plus subscriptions and API spend, I usually land somewhere between $1,000 and $2,000 a month.
Taken in isolation, that is a lot of money. I still remember when people were seriously debating whether ChatGPT was worth $20 a month. We were young. We were innocent.
Now I am happy to spend a lot more than that.
With agents, just like with people, the part that really matters is the brain. The brain is the primary model you give it. It reads for you, reasons, calls skills and workflows, delegates to sub-agents, pulls their work back in, arbitrates, and synthesizes.
When you hire, you try to attract the best people you can. So when it comes to your personal agent, probably your closest collaborator, the one with access to a large part of your private and professional life, are you really going to cheap out and give it a stupid brain?
Oslo, my OpenClaw agent, named after my first dog when I was a kid, runs on the best model available all the time, for everything.
Most of the time, with the highest reasoning setting too.
I do not care what one prompt costs me. I care about the quality of the work and the volume of value I can produce with it.
A mediocre output is not cheaper if it has to be corrected, reframed, re-contextualized, or thrown away. I am not even getting into the damage a weak model can do when it improvises badly, or when it lets itself get hit by prompt injection because it is simply not trained to resist it.
Before I come back to the money, a bit of context.
I installed my first OpenClaw agent in December. So we are talking about less than 4 months, and subjectively something like 10 years.
My first Clawdbot was a pain in the neck to manage and stabilize. Total chaos, daily memory loss, Gateways refusing to come up, the usual fun. But even in that mess, I knew I had never seen anything more powerful or more useful.
Back then, and by back then I mean 4 months ago, Oslo could be dumb as a brick one minute, then 5 minutes later give me 2 or 3 prompts that felt like touching AGI with my fingertips.
If you play golf, you know the feeling. You hit 20 shots like a man who has never held a club in his life. You are one swing away from throwing the whole bag into the water. Then one almost perfect shot comes out of nowhere, and suddenly you are ready to suffer some more.
Oslo has put me through all of it. And yes, I have insulted him more than I have ever insulted any person or object in my life. Still, after all this time, and yes, 4 months, we are still together.
Since those messy beginnings, OpenClaw has moved incredibly fast. If you have been paying attention, you know that already. What I first dropped on my VPS and what I have today are two different worlds.
For context, here is my current setup.
I have gone almost all in on OpenAI.
Oslo runs on GPT 5.4 Think High as the default model, GPT 5.4 xHigh for sub-agents with no time limit, and Codex sessions through ACP only, also on GPT 5.4 xHigh, in Full Access.
If I had stayed on Anthropic, I would probably be on Opus 4.6 Think High as the default model, Opus 4.6 Think High for sub-agents too, and Claude Code sessions through ACP in YOLO mode.
A slightly more technical note: today’s release, March 8, 2026, made this setup much easier, especially with the progress on ACP and Discord channel support for Codex and Claude Code sessions.
A slightly less technical note: yes, use Discord, create one channel per major topic, and thank me later.
I am not claiming this is universal truth, and I am not saying everyone should copy me exactly, except for Discord. I am saying this is what works best for me today.
One thing matters much more than it looks, and I did not understand it right away: the need for consistency between the main model and the sub-agents.
Using the same model family for the default model and its sub-agents is critical.
You can ignore that and the system will still work. But you will create friction all along the chain from prompt to output. Skills are not written the same way. Memory files are not structured the same way. Even when context gets passed along, part of the cognitive continuity disappears when the models do not think with the same reflexes.
You keep the history. You keep less of the working logic.
At the end of the chain, your agent still gets a result. It also gets more noise, more mismatch, and lower quality than it was actually capable of producing.
That is also why, because I use OpenAI for the main model, I go all the way to Codex for the ACP runtime. If you prefer to use Anthropic, same logic with Claude Code. Same family, same instincts, less friction, better handoffs.
And yes, all of this in High or xHigh.
It is slower. It is more expensive.
Honestly, I do not care.
I would rather wait a bit longer and get solid work back than save 30 seconds and lose 20 minutes repairing a broken output, a weak analysis I cannot trust, or a sub-agent that wandered off into the weeds.
That does not mean you should burn tokens like a maniac. It means your default setup should usually bias toward the highest level of intelligence you can justify.
Then, when a task clearly does not need that much brainpower, lower it case by case. Heartbeat is a good example. It runs simple checks and simple reporting. It does not need the same brain as research, writing, analysis, coding, or orchestration.
Back to the money.
Spending $1,500 a month can sound high. Fair enough. If your agent is mostly checking whether you got an email and reminding you what is on your calendar, then yes, that is expensive.
But OpenClaw is not for that.
We all have different uses. Here is mine.
Oslo helps me manage everything I do across every company where I have a role. He does deep research, writes code like a psychopath, builds websites and presentations in minutes, manages our RAG agents and their Source of Truth from a casual remark, generates individual reports for each Nimbus Territory Manager before the Monday sales meeting, with recommendations on pipeline priorities, manages some of our Meta campaigns, and does dozens of other things that create very high value for me.
And yes, he also tells me when an important email comes in or what is on my calendar. But that is just because if a system can do the hard things, the easy ones come for free.
If I try to put that output in perspective using US market rates, I am probably looking at something like $25,000 to $30,000 a month in salaries. Executive assistant, developer, designer, data analyst, marketing manager. The comparison has limits, of course. But it is good enough to put the church back in the middle of the village.
Oslo is not free, but today he is, by far, my best investment.
Take that from someone who bought Bitcoin in 2013.
He is not replacing a team I would otherwise have hired. I would not have hired that team. Too much nice to have for a $30,000 monthly burn, not enough immediate must have.
He is increasing me. Multiplying me.
He lets me do better and faster work, including things I could not have done myself anyway, and create value that simply would not have existed otherwise.
So if you think SOTA model tokens are too expensive, you have 2 options.
Lower your expectations and accept the consequences of working with weaker models, though not completely stupid ones, something like Gemini 3.1 Pro for example. Patience is on your side. The price of intelligence will keep dropping.
Or create more value with what you are doing, until the cost stops being the story because the upside is so much larger.
The price of anything is always relative. You do not need a plane to go buy bread. For crossing the Atlantic, it is still useful.
The market still underestimates how cheap AI is for what it actually makes possible, and OpenClaw in particular.
I do not see it as a slightly expensive software line on a P&L. I see it as an investment with the power to multiply the number at the bottom right of that same table.
So no, Oslo is definitely not free.
And I have never been happier to spend this much money on 0s and 1s.
P.S. For the price of a Mac Mini you probably did not need to buy, you can get yourself a serious VPS and 6 months of ChatGPT Pro. Who knows what you could build in 6 months.