Cost Control:
Your AI Workforce Budget

The number: what an AI workforce actually costs

Let's start with the number your CFO will ask about. A mid-size deployment — 7 event-driven agents across 5 departments, handling ~265 tasks per month — costs $210/month. That includes compute ($31) and model inference ($180).

Scale it to 50 agents running 4 hours/day each and the total is $1,320/month. The compute (sandbox VMs, browser containers) is $120. The model inference (Claude Sonnet API calls) is $1,200.

The human equivalent of this work — QA testing, code review, competitive intelligence, accounts payable, data processing, desktop automation — would cost $25,000–120,000/month in fully-loaded salary. The AI workforce costs 1–5%.

Two line items, not one

Every AI agent bill has two components. Understanding which one to optimize is the difference between saving pennies and saving dollars.

Compute: the machine the agent runs on ($0.001–0.50/task)

This is the sandbox, the browser container, the GPU instance. It's metered by the second. A ab0t.micro running for 5 minutes costs $0.001. A GPU instance running for 20 minutes costs $0.18. The platform tracks this automatically per sandbox.

Compute is 10–20% of total costs. It's the smaller line item. Optimizing compute (smaller instances, shorter runtimes) saves real money at scale, but it's not where the big dollars are.

Model inference: the brain ($0.01–5.00/task)

This is the Claude or GPT API bill. Every screenshot the agent reasons about, every code review it writes, every data extraction it performs costs tokens. A simple task (extract pricing from a webpage) might cost $0.01 in tokens. A complex task (full code review with visual regression testing) might cost $0.50.

Model inference is 80–90% of total costs. This is where the money goes. And it's where the biggest optimization lever is: model selection. Claude Sonnet 4.6 costs ~3x less than Opus 4.6 per token and handles 95% of agent tasks at equivalent quality.

What each department actually spends

Based on the agent fleet described across this guide series, here's what a typical monthly bill looks like broken down by department:

Engineering is the biggest spender because it has the highest task volume (every PR, every deploy). But it's also generating the most value — $10,800/month of QA and code review time replaced for $97. The ROI is immediate and obvious.

Four controls that prevent surprises

AI costs are predictable when you have the right controls. Without them, a misconfigured agent or a forgotten sandbox can create an unexpected bill. Here are the four controls every deployment needs:

1. Auto-stop (prevent runaway compute)

Every sandbox and container has an idle timeout. When the agent finishes and goes idle, the sandbox stops automatically. Billing pauses. Files persist. This is the single most important cost control. Set it on every sandbox, no exceptions.

The horror story: A developer creates a GPU sandbox (ab0t.gpu, $0.526/hour) for a 20-minute ML inference job. They forget to stop it. It runs for a month. Cost: $378 for a task that should have cost $0.18. Auto-stop at 30 minutes would have capped it at $0.26.

2. Spending caps (prevent budget overruns)

Set monthly limits per user or per department. When the limit is reached, new sandboxes can't be created until the next billing cycle. This is the hard guardrail that prevents any single team or person from blowing the budget.

3. Per-task metadata (enable accountability)

Tag every sandbox with department, task type, and cost center. Without tags, you have a single bill with no visibility into who spent what. With tags, you can produce the departmental breakdown that finance needs for budget allocation.

4. Monthly reporting (maintain visibility)

Pull a monthly cost report that shows spending by department, by agent, and by task type. Compare to human-equivalent costs. Identify agents that cost more than expected or deliver less value than projected. The report takes 5 minutes to generate from the platform's cost API.

The five most common cost mistakes

1. Using Opus when Sonnet works fine

Claude Opus 4.6 is 3x more expensive than Sonnet 4.6 per token. For most agent tasks — code review, data extraction, vision analysis, report generation — Sonnet produces equivalent results. Reserve Opus for genuine deep-reasoning tasks: architectural decisions, complex multi-step planning, nuanced judgment calls. Default to Sonnet everywhere else.

2. Sending full-page screenshots when you only need a section

Each vision API call costs based on image size. A full-page screenshot of a long pricing page might be 4000x8000 pixels. Crop to just the pricing table before sending to the model. Or extract the text first and only use vision when text extraction fails.

3. Running ab0t.medium for tasks that fit on ab0t.micro

ab0t.medium costs 4x more than ab0t.micro. Most orchestration tasks, API calls, and light data processing run fine on 1 GB RAM. Only upgrade when you hit a concrete memory limit. The default should always be the smallest instance.

4. Forgetting auto-stop on GPU instances

A ab0t.gpu costs $12.62/day. Forgetting to stop it for a weekend costs $25.24 for zero value. GPU instances should always have auto_stop_minutes: 30 at most. The inference job finishes in 20 minutes; there's no reason for the GPU to idle.

5. Not tracking model costs separately from compute

The platform tracks compute costs automatically. Model costs come from your Anthropic/OpenAI billing. If you only look at the platform bill, you're seeing 10–20% of the true cost. Combine both into a single report for an accurate picture.

How costs scale: from 7 agents to 50

Costs scale linearly with the number of tasks, not the number of agents. Adding an agent that runs once a week costs ~$2/month. Adding an agent that runs 100 times a day costs ~$65/month. The cost is proportional to the work done, not the infrastructure provisioned.

At every scale, the total AI workforce cost is a fraction of a single human employee's salary. The question isn't "can we afford AI agents?" It's "how many should we have?"

The cost trend: it only gets cheaper

Model inference prices have dropped 10x in the last 18 months. Claude Sonnet's cost per token in early 2025 was 5x what it is in March 2026. GPT-4o's pricing followed the same curve. The trend is clear and accelerating.

At the current rate, model costs will drop another 3–5x by early 2027. A deployment that costs $1,320/month today will cost $300–450/month in a year, for the same work. The ROI improves every quarter without any action on your part.

Compute costs are also declining (AWS price reductions, Graviton instances, spot pricing) but more slowly — ~10–15% per year. Since compute is only 10–20% of the bill, this matters less.

The implication: the cost case for AI agents gets stronger with time, not weaker. Early adopters pay more per task than late adopters will, but they capture the productivity gains now. The late adopter saves $1,000/month on model costs but misses 12 months of 98% labor savings.

The monthly report template

Here's the report format that finance teams have told us works. One page, five sections, all the numbers.

What's next