CI baseline: what CI costs and how fast it ships, before the changes

The "before" numbers to compare against once the app is split and CI skips tests a PR can't affect.

Nov 1, 2026 · Oct 18–29 (before Fede's Oct 30 changes) · re-run with bin/ci-baseline

About 84% of the AWS bill is CI. Every CI machine carries a label AWS bills by, so this is measured, not guessed: $5,117 of October's $6,066 in usage before credits. Nearly all of October's $678 credit landed on CI machines, so after credits it is still about 83%.

Number	Before
CI spend per day	$100
CI spend per merged PR	$0.69
PRs merged per day	145
CI runs per merged PR	5.6
Merge → live (median)	54 min
Open → live (median)	123 min
PRs that touch no product code	11%

What this says about the plan. Splitting off the website would skip almost no CI: no PR in two weeks touched only website files. Tools, evals and docs-only PRs are about 9%, so path-scoped CI can skip at most about a tenth of PRs. The bigger levers are repeat runs (5.6 CI runs for every merged PR: 3.7 on the PR, 1.1 in the merge queue, 0.8 on main after) and a few slow steps.

The rest of the bill: the $411 "on-demand" line is $219 of full-price CI machines (used when cheap spare ones weren't available) plus $192 of data sent out of AWS; the $243 "VPC" line is the public IP address each runner gets. Vercel's bill can't be read from our login.

Day by day, top time sinks, sources and method

Per day (UTC)

AWS dollars are usage before credits. "CI $" = everything tagged Purpose=ci. Job-hours are estimated (runs × median machine-minutes per run, from 30 sampled runs of each kind). Grey rows are after Fede's Oct 30 04:14 UTC changes; Oct 31 is still filling in at AWS.

Day	Spot EC2 $	On-demand EC2 $	CI $	PR runs	Queue runs	Main runs	Job-hrs	Merges	CI $/merge
10-18	13	2	21	79	50	43	86	45	$0.47
10-19	59	6	83	495	140	111	412	134	$0.62
10-20	89	8	122	619	204	157	535	184	$0.66
10-21	33	4	49	232	72	57	197	69	$0.71
10-22	90	7	121	640	183	123	531	165	$0.73
10-23	61	10	89	512	137	121	422	137	$0.65
10-24	41	8	64	395	97	78	318	98	$0.65
10-25	44	7	68	346	93	83	285	94	$0.72
10-26	66	10	94	451	165	81	395	125	$0.75
10-27	115	23	171	866	231	188	711	219	$0.78
10-28	117	22	172	987	298	210	831	263	$0.65
10-29	99	16	143	777	234	173	655	211	$0.68
10-30	133	33	202	1023	318	203	865	270	$0.75
10-31	70	10	101	538	205	153	484	183	$0.55
Total	1031	167	1500	7960	2427	1781	6724	2197	$0.68

Oct 18–29 only: CI $1,197 ($100/day), spot $827 vs on-demand $123, 1,744 merges, 9,728 runs (3.7 PR + 1.1 queue + 0.8 main per merge), ~448 machine-hours a day (estimate, ±30%).

Speed (PRs merged Oct 25–31)

Window	PRs	Open→merge	Merge→live med / p90	Open→live med / p90	Live within 1 h
Before Oct 30 04:14Z	942	66	54 / 77	123 / 328	2%
After (2 days)	423	69	22 / 39	95 / 348	23%

Minutes. The "before" row matches Fede's own baseline for the week of Oct 26 (125 / 54 / 2%), which checks the method. The "after" row is two days and already shows his changes working.

What merged PRs touch (Oct 18–31, 2,197 PRs)

Only touches	PRs	Share
Website (src/app/(public), components/marketing, config/site-paths.json)	0	0%
Internal tools (admin/dev, agents, api/agents, api/admin/dev, components/maestro)	6	0.3%
Tools (scripts, evals, docs)	198	9.0%
CI plumbing (.github, scripts/ci)	10	0.5%
A mix of the four above, no product code	20	0.9%
Product code plus one of the above	936	42.6%
Product code only	1,027	46.7%

11 commits touched website files in the window, all alongside product code. PRs with 100+ files, or whose file list could not be read, count as mixed.

Where CI time goes (90 sampled runs, 30 per kind, before Oct 30)

Median machine-minutes per finished run: PR 38, merge queue 36, main 8. These are the noisiest numbers here (a 30-run sample moved the PR median between 20 and 38), so job-hours above are ±30%. Compare the per-run step medians below, not totals.

Step	Median min per run that ran it	Runs (of 30)
PR · Unit Tests: run this shard	19.6	16
Queue · Full set: full unit tests	28.8	7
Queue · Full set: precise unit tests	9.6	9
Queue · fail-fast watcher (a small machine that waits out the whole queue run)	8.9	30
PR · Unit Tests: git checkout (shallow since Oct 31)	4.3	30
Queue · Guard pins	4.1	30
Queue · affected unit tests	3.2	30
Main · Preflight wait-then-cancel	3.1	30
PR · CI Checks: Mutation controls	2.0	29

Mutation controls is the step to watch: 2.0 minutes on a typical PR run but 5.4 on average and up to 31, so a few PRs pay most of it (158 machine-minutes across 29 runs).

The AWS lines, explained

CI share. Over Oct 18–31 alone it is 78% of usage ($1,500 of $1,912); October as a whole is 84%. The Purpose tag is switched on for billing, and Purpose=ci covers every EC2 machine-hour ($4,220.54 tagged = $4,220.54 total in October), plus CI's IPs, disks, logs, queue and API. October: $5,117 of $6,066 usage. Fede's "$75/day" counts only spend above the Sep 22–28 level; this page counts all of it, hence $100.
Which workflow. "CI" here is everything our CI machines run, including evals and replays, not only ci.yml. To split by workflow, switch on the runs-on-workflow-name billing tag (exists, currently inactive; counts from the day it is switched on).
On-demand $411 (October). $219 is BoxUsage on CI runner types (r8g.xlarge $65, t4g.medium $53, m7g/m8g.2xlarge), all tagged CI: spot fallback plus queue checks deliberately on on-demand. $192 is data transfer out, which AWS files under the same purchase type. The daily on-demand column includes both.
VPC $243. PublicIPv4:InUseAddress ($221 us-east-1, $22 us-west-2). No NAT gateway charges.
Vercel. The CLI (logged in as g-era, team PropFlowAI) answers "Costs not found (404)"; billing needs the dashboard.

Sources

AWS Cost Explorer daily, UnblendedCost, Usage records, pulled Nov 1. GitHub: every ci.yml run by event in 6-hour windows; merged PRs per 2-hour window with file lists; Production deployments by vercel[bot]. Live = first Production deployment containing the merge commit (Fede's definition, ci-two-hours). Re-run: propflow-docs/bin/ci-baseline --from 2026-10-18 --to 2026-10-31 --split-at 2026-10-30T04:14:00Z.