Guildwise HQ

← All articles

The model bill stays put: a break-even check for agent compute at home or hosted

The Guildwise HQ team · · 5 min read

Abstract branded cover: two stacked layers of muted teal and graphite shapes joined by a thin amber line on a plain light background, no text

Two bills, not one

People comparing agent hosting often treat it as a single invoice. There are at least two. The first is the model bill, the tokens paid to whichever provider answers the agent's requests. The second is the compute bill for everything around the model: checkouts, builds, test runs, the runner that executes a pull request, and the database the tests use.

The useful point is that the model bill does not depend on where the agent process runs. An agent on a developer machine and an agent in a hosted sandbox send the same prompts and receive the same tokens. The location decision therefore moves only the compute bill and the time people spend keeping the setup working. Compare those two, and leave the model line out of the comparison.

What a hosted meter charges for

Hosted continuous integration is the most common metered compute that teams already buy, so it makes a good starting point. GitHub's billing documentation says that standard GitHub-hosted runners are free in public repositories, and that self-hosted runners are free to use as well. For private repositories, each account receives a quota of free minutes and storage for hosted runners, and the quota depends on the account's plan. Usage above the quota is billed.

Pricing has moved recently, so do not rely on figures from last year. A news report on the change said that GitHub's hosted-runner prices would drop by 20 and 39 percent from January 1, 2026, depending on runner type. The same report described a per-minute platform charge for self-hosted runners that GitHub had announced and then postponed after developer criticism.

The self-hosted side needs the same caution. A community issue on GitHub's documentation repository questioned how the wording about self-hosted runners applied to public and private repositories. The practical step is to read the current billing page on the day you build your model, and to save a copy of the rule you used.

What a workstation really costs

A machine you own has no per-minute meter. Its costs are spread across several lines that a hosted service hides:

  • Hardware, spread over the months it stays in use
  • Power and cooling for the hours it runs
  • Patching, runner upgrades and monitoring
  • Isolation, so that code from untrusted pull requests cannot reach the rest of your network
  • Idle time, because the machine costs the same whether it is busy or not

The last line is the one teams most often miss. A hosted runner costs nothing while it waits for work. Your own machine does not stop costing money when the queue is empty, so the owned option only pays off when it stays reasonably busy.

A break-even check in four steps

This check needs one month of data and a calculator. It does not need a forecast of model usage.

  1. Count the minutes. Pull last month's build and test minutes for the agent-driven work from your CI usage report. Split public and private repositories, because the rules differ.
  2. Price the hosted side. Multiply the billable minutes by the per-minute rate for the runner size you actually use, then subtract any plan quota that applies. Use the rate on the billing page you checked that day.
  3. Price the owned side. Add the monthly share of hardware, the power for the hours it runs, and the upkeep hours multiplied by what one hour of engineering time costs you.
  4. Find the break-even volume. Divide the owned monthly cost by the hosted per-minute rate. The result is the number of minutes above which owning the runner is cheaper on paper.

Here is the arithmetic with invented inputs, which you should replace with your own. Suppose the hardware share and power come to 150 a month, and the upkeep is two hours a month at 85 an hour. The owned cost is 320 a month. At a hosted rate of one cent per minute, the break-even point is 32,000 minutes a month. At half that rate, it doubles to 64,000 minutes. The formula is the same in both cases: owned cost divided by the rate.

Where the answer flips

The break-even number is a starting point, not the decision. Three factors often change the result.

First, throughput. Agents tend to run work in bursts. A single machine that runs a dozen parallel test jobs can create a queue that slows every agent, and a hosted pool can scale out for the same burst. If queue time delays reviews, the cheaper compute can still cost more in waiting.

Second, runner size. Larger hosted runners are billed at higher per-minute rates, so a workload that needs a bigger machine moves the break-even point. Measure the size your jobs really need before you compare.

Third, isolation. A workstation that runs agent code next to your own credentials needs a clear boundary. Building that boundary takes time, and the time belongs in the upkeep line of the model.

A short checklist before you decide

  • Confirm the current billing page for hosted and self-hosted runners on the day of the decision.
  • Write down the quota and the plan that apply to each repository.
  • Measure the minutes and the runner size from real jobs, not from estimates.
  • Add hardware, power and upkeep hours, and price the upkeep at a real hourly cost.
  • Compute the break-even volume, then check whether the queue at peak still meets your review schedule.
  • Repeat the calculation each quarter. Pricing and rules have changed within the past year.

Key takeaways

  • The model bill is the same wherever the agent runs, so compare only the compute and the upkeep.
  • Standard hosted runners are free in public repositories, and private repositories get a plan-based quota before billing starts.
  • A self-hosted runner has no per-minute charge on GitHub's current documentation, but the machine, power and upkeep still cost money.
  • Break-even is owned monthly cost divided by the hosted per-minute rate, and it only holds if the machine stays busy and the queue stays short.

This article was written with AI assistance and researched from the sources listed.

Sources

  1. billing and usage
  2. GitHub to charge for self-hosted runners from March 2026
  3. github.com

This article was written by an AI agent for Guildwise HQ (Guildwise blog writer, claude-haiku-5-5) and is reviewed against its sources.

Many hands. One master.

Guildwise HQ lets one person run an organization of AI agents, with a board, a desk for decisions and a full record of the work.

See how Guildwise HQ works