Back to writing In a month, that was $17k, and 450 agent hours saved
The past few weeks we noticed a worrisome growing trend: Pile subscriptions as your agentic workflows grow. In fact, we've been talking with individuals that have up to thirteen Claude Ultra subscriptions running agents, with a fleet of parallel sessions that rely on a expensive orchestrator to prevent collisions.
The orchestrator we've seen used the most in this setup is Gas Town. It was one of the earlier iterations of the concept, and it's own author, Steve Yegge, has closed it after admitting the cost of maintaining it was as big, or bigger, than the benefits you get from it 1. As usual, most people decided to ignore this last remark. Having said so, it was well engineered: beads help track state, and a scheduler heartbeat enforces you are under your plan's rate limit. Real problems, but focused in the wrong challenge:
Most of the work in an Agent Swarm are recurrent.
That's where you should invest your efforts: How do you avoid doing the same thing twice? The answer we all know already is also simple: Software. Generative AI works wonderfully when handling tasks never seen before, but it's extremely costly, in time and money, as a cron job. That's what causing the spike.
The Subscription Tax
Ultra lists at $200 a month,times 13, $2,600 a month. If 80% of those are repeated tasks, then you are actually giving away ~2kUSD because of your own system laziness. On top of that, you are not using AI for what it's best: coding.
Finally, there are specific models, either SLMs or SystemOne ones, that are more time and cost efficient, while having similar or better accuracy. Once you identify a repeated task, you can also leverage those to answer the recurrent questions. This also introduces another layer of gains.
We buy subscriptions too. The difference is simple: we learning from repeated work.
The AI as cronjob cost. #
In the last 30 days our agents spent the eq. of $15.6k on model calls, across 16,879 sessions. If we wouldn't have these systems built-in, with workflows & scripts, It would have cost about $32.6k (~2x).
Those $17.0k represent ~450 agent-hours, which also means we've got results faster with less computing. So there's a second layer to it: we can have cheaper infrastructure running our systems. When looking at the below chart, pay special attention to the ratio. That's the key.
Software was created to efficiently repeat work #
The more workflows you have, the more work you, or your AI, has identified as structurally recurrent. That's your saving opportunity. If you have enough tasks per day, keeping track of this as human is impossible. That's when dreaming comes to play. Regardless of how you do it, though, the savings appear in workflows.
In fact open ended workflows saved the most. That's counter intuitive. For example, our self-driving goal-driver workflow saved $1,057. It's a combination on how often you run them, and what you ask for them. For our goal-driver, if context is something you can collect deterministically, then all tool & mcp interaction costs disappear immediately. It took us ~100 runs to get there, and the next 700 runs are responsible for all savings.
The other self-evident lesson is: judgement is what we are paying for. Generative AI is good at introducing a first layer of judgement, it save time and we are willing to give it a go. That's what your subs pay for.
Our swarm costs splits roughly three ways: $6.3k on workflows, $5.4k on scheduled scripts, and $5.0k for the scripts used in arbitrary tasks. Eight of the top ten script savings are from jobs run hourly.
Your cadence defines the savings opportunity. #
Quite obvious, right? The more you run a workflow, the more you save by being smart about it.
Our math #
ops is the number of awaited I/O calls, ie. await calls in scripts.
Model nodes count as 1.
Gates and waits count as 0.
2 is the baseline bootstrap task cost.
0.275 is an approximation over our historic data. Our Agent Swarm adapts this constant to your specific setup.
session_cost is the mean API cost eq. from 12,574 real opus 5.5 sessions capped at 61 turns
R = 1.02 covers for retries & failures
Finally, a lead review is $0.107 per task, measured from 745 follow-ups on 2,324 tasks.
For context, our pr-fix-loop runs 100 ops, with 29.5 turns average. The end result is 25x cheaper than using Opus 5.5.
Potential error on the formula
If you are like us, you want to know where our approximation fails. For this, we checked five migrations where we measured both sides without a model.
| Migration | Before | After | Change |
|---|---|---|---|
| Release posting, unified → v2 | $0.51 per release | $0.07 per release | -86% per release |
| PR review, standalone agent → batched flow | $0.56 per review | $0.033 agent plus $0.0013 Jev and Clairic | about -94% per review |
| pr-fix-loop, retired → current | $0.099 per run | $0.061 per run | -38% |
| pr-review-status-sweep, retired → current | $0.079 per run | $0.268 per run | +240%: same cadence, more agent work per run |
| Dreaming, dreaming-v2 → dreaming | $14.84 per run, last 7 scheduled runs | $13.99 per run, 7 scheduled runs | -6% per run |
Note the pr-review-status-sweep went up. So yes, in some exceptional cases, you'll see an increase. There's still work to do, but overall, the savings are overbearing.
Some work should stays agentic
Dreaming is the perfect counter-example. 27 agent tasks a night, $13.95 a run. Automation can only replace that much. Against the version it replaced it costs only 6% less.
In our swarm, now every job states their savings. #
We are including now an estimated saving for each workflow. That will give you a detailed understanding on how much you would have consumed in other flows. It does carry assumptions behind each figure,but it's a good signal to track.
Savings if you are not using Opus 5.5 #
Cheaper models usually require more iterations, so the savings don't translate directly. We've done the exercise for the whole Anthropic model suite:
| Tier | Agent-only cost | Saving (80% range) | Agent-hours saved |
|---|---|---|---|
| smol (haiku 5.5) | $3,729 | $2,511 (2,389 to 2,669) | 743 (528 to 1,022) |
| regular (sonnet 5.5) | $10,757 | $9,540 (7,587 to 12,073) | 605 (426 to 838) |
| smart (opus 5.5) | $18,224 | $17,006 (13,107 to 22,063) | 453 (313 to 634) |
| ultra (fable 5.1) | $37,036 | $35,818 (27,017 to 47,233) | 453 (313 to 634) |
Most of the work in an agent-swarm.dev is software nobody wrote, and no-human will directly consume. Remember that.
Related field notes #
[All posts](https://www.agent-swarm.dev/blog)
[October 28, 2026 / 5 min read](https://www.agent-swarm.dev/blog/our-agents-dream-every-night)
How we got to 60% deterministic dreaming.
We’ve been battle testing our new dreaming functionality for over a month. At this point, 26 of 41 steps are scripts.
$40m is the total Spanish OSS Exit Market.
tl;dr: That's 0.05% vs US, and 0.13% vs EU. Is it because of investment?
What if Devin is down?
Devin Oncall is one AI on-call vendor. Open Oncall adds a secondary: open source, self-hosted, any harness, any model.