Start with the $0.002. GitHub introduced a per-minute Actions cloud platform charge that applies to every Actions workflow, and on GitHub-hosted runners it is folded into the reduced meter rate rather than billed as its own line [6]. The self-hosted version of the same charge would have been itemised against minutes on machines you already operate [9]. Same number, absorbed in one place and visible in the other. Standard runner usage on public repositories stays free either way [5].
The cuts are not proportional. That changes matrix design. A Linux 64-core arm64 minute used to cost 20 times a Linux 2-core minute, at $0.160 against $0.008. It now costs 16.3 times, at $0.098 against $0.006 [2][2]. Wide machines took the deeper reduction, so consolidating a sharded suite onto one big runner is cheaper against its alternative than it was in December. In absolute terms that box is $5.88 an hour, down from $9.60 [1].
The escape hatch is where the arithmetic bites. If the self-hosted charge returns as written, a 2-core Linux runner in a private repo has to run all-in for less than $0.004 a minute, or $0.24 an hour, to stay ahead of the hosted rate [3]. That budget covers hardware, power, image maintenance, and whoever restarts the wedged agent at 2am. At the top of the range the same charge is noise: a 64-core arm64 equivalent has $0.096 a minute of headroom, $5.76 an hour, before hosted wins [4]. The proposed fee hits small self-hosted runners hardest, pricing them out to save a fraction of a cent a minute, which describes a lot of the fleets whose owners objected; it does not price out self-hosting in general.
Per SamExpert's reporting, developers called billing for hardware they already own "absolutely bananas" and compared it to paying rent on their own property [11]. GitHub said it "missed the mark with this change by not including more of you in our planning" [13]. It had framed the charge as ending a cross-subsidy in which hosted-runner revenue underwrote the cost of operating Actions for everyone [10].
The reassurance figure was that 96% of customers would see no change to their bill, and that within the affected 4%, 85% would go down while the remaining 15% faced a median increase of about $13 [8]. That last group is 0.6% of customers [5]. An aggregate like that says nothing about your invoice, which is why GitHub pointed at its pricing docs and calculator rather than publishing a single headline rate [15].
None of this touches the other half of the meter. Cheaper minutes lower the per-minute floor and do nothing for the minutes burned on flaky failures and re-runs, or the engineer hours spent reading a log to learn that a disk filled up [16]. Roughly 73% of organizations report no use of AI in their CI/CD pipelines at all [18], so the volume arriving from AI-assisted authoring meets a pipeline that is cheaper per minute and no better per failure.
