GitHub Actions Runner Build Benchmarks: Rust, TypeScript, Docker

At 4 vCPU, RunsOn is the best value on Rust and Docker, and ties with boat on TypeScript.

Rust spotify-player

RunsOn c8a.xlarge · EBS gp3

2m 02s $0.0036 per build

GitHub 4-core: 13× the cost, 1.9× the time, 9th of 9

TypeScript opencode

boat 4 vCPU ≈ RunsOn r8a.xlarge · tmpfs

1m 36s $0.0010 per build

GitHub 4-core: 26× the cost, 1.3× the time, 8th of 8

Docker PostHog image

RunsOn r8a.xlarge · tmpfs

5m 15s $0.0087 per build

GitHub 4-core: 20× the cost, 2.8× the time, 7th of 7

Best value per provider

Each provider's best variantEach provider's best 2 vCPU variantEach provider's best 4 vCPU variantEach provider's best 8 vCPU variant, ranked by p90 time × cost per build.; shades are build phases in order.; build time on the right.

Rustspotify-player

RunsOn4 vCPU · 8 GBc8a.xlarge · EBS gp32m 02s, cost per build spot $0.0036, on-demand $0.0087, best-value of 9 machines; phases: setup 15s, cargo test 1m 01s, clippy 34s, runner overhead 12s, ×1.00
Avrea4 vCPU · 16 GB1m 37s, cost per build $0.0129; phases: setup 13s, cargo test 49s, clippy 25s, runner overhead 10s, ×2.98
Namespace4 vCPU · 8 GB4x81m 46s, cost per build $0.0106, best-value of 2 machines; phases: setup 10s, cargo test 52s, clippy 29s, runner overhead 15s, ×3.16
Ubicloud4 vCPU · 16 GBstandard2m 38s, cost per build $0.0066, best-value of 2 machines; phases: setup 23s, cargo test 1m 18s, clippy 42s, runner overhead 15s, ×4.42
boat4 vCPU · 8 GB4m 48s, cost per build $0.0029; phases: setup 48s, cargo test 2m 16s, clippy 1m 06s, runner overhead 39s, ×5.41
Warpbuild4 vCPU · 16 GB2m 23s, cost per build $0.0191; phases: setup 12s, cargo test 1m 22s, clippy 43s, runner overhead 6s, ×7.41
Blacksmith4 vCPU · 16 GB1m 59s, cost per build $0.0159; phases: setup 17s, cargo test 57s, clippy 32s, runner overhead 13s, ×18
AWS CodeBuild4 vCPU · 8 GBmedium3m 48s, cost per build $0.0380; phases: setup 32s, cargo test 1m 55s, clippy 1m 04s, runner overhead 17s, ×20
GitHub4 vCPU · 16 GB3m 48s, cost per build $0.0456; phases: setup 21s, cargo test 2m 05s, clippy 1m 10s, runner overhead 12s, ×31

Bar shades, light to dark, in execution order: setup = checkout + packages + toolchaincargo fmtcargo testclippyrunner overhead

cost per build× = p90 time × cost against the best

TypeScriptopencode

boat4 vCPU · 8 GB1m 36s, cost per build $0.0010; phases: toolchain 1s, clone 4s, bun install 14s, typecheck 1m 00s, runner overhead 17s, ×1.00
RunsOn4 vCPU · 32 GBr8a.xlarge · tmpfs1m 03s, cost per build spot $0.0021, on-demand $0.0074, best-value of 9 machines; phases: toolchain 1s, clone 1s, bun install 9s, typecheck 38s, runner overhead 15s, ×1.02
Avrea4 vCPU · 16 GB50s, cost per build $0.0067; phases: toolchain 1s, clone 3s, bun install 5s, typecheck 30s, runner overhead 11s, ×2.60
Ubicloud4 vCPU · 16 GBstandard1m 52s, cost per build $0.0047, best-value of 2 machines; phases: toolchain 2s, clone 5s, bun install 25s, typecheck 1m 01s, runner overhead 21s, ×4.18
Namespace4 vCPU · 16 GB4x1653s, cost per build $0.0106, best-value of 2 machines; phases: toolchain 1s, clone 1s, bun install 11s, typecheck 33s, runner overhead 8s, ×6.02
Warpbuild4 vCPU · 16 GB1m 21s, cost per build $0.0108; phases: toolchain 1s, clone 1s, bun install 14s, typecheck 56s, runner overhead 10s, ×7.21
Blacksmith4 vCPU · 16 GB1m 45s, cost per build $0.0139; phases: toolchain 1s, clone 2s, bun install 12s, typecheck 1m 05s, runner overhead 25s, ×14
GitHub4 vCPU · 16 GB2m 05s, cost per build $0.0250; phases: toolchain 1s, clone 2s, bun install 24s, typecheck 1m 11s, runner overhead 28s, ×32
AWS CodeBuildfailed · out of memory

Bar shades, light to dark, in execution order: toolchainclonebun installtypecheckrunner overhead

cost per build× = p90 time × cost against the best

DockerPostHog image

RunsOn4 vCPU · 32 GBr8a.xlarge · tmpfs5m 15s, cost per build spot $0.0087, on-demand $0.0307, best-value of 8 machines; phases: fetch source 5s, buildx setup 3s, docker build 4m 42s, runner overhead 25s, ×1.00
Avrea4 vCPU · 16 GB5m 26s, cost per build $0.0435; phases: fetch source 23s, buildx setup 3s, docker build 4m 32s, runner overhead 28s, ×3.43
Namespace4 vCPU · 16 GB4x165m 02s, cost per build $0.0604; phases: fetch source 20s, buildx setup 2s, docker build 4m 26s, runner overhead 14s, ×6.86
Blacksmith4 vCPU · 16 GB5m 53s, cost per build $0.0471; phases: fetch source 20s, buildx setup 2s, docker build 5m 01s, runner overhead 30s, ×7.18
Warpbuild4 vCPU · 16 GB7m 29s, cost per build $0.0599; phases: fetch source 20s, buildx setup 4s, docker build 6m 48s, runner overhead 17s, ×7.76
Ubicloud4 vCPU · 16 GBpremium11m 01s, cost per build $0.0441, best-value of 2 machines; phases: fetch source 20s, buildx setup 9s, docker build 9m 46s, runner overhead 46s, ×12
GitHub4 vCPU · 16 GB14m 37s, cost per build $0.1754; phases: fetch source 24s, buildx setup 7s, docker build 12m 22s, runner overhead 1m 44s, ×42
AWS CodeBuildnot run: needs more than 8 GB of memory
boatnot run: needs more than 8 GB of memory

Bar shades, light to dark, in execution order: fetch sourcebuildx setupdocker buildrunner overhead

cost per build× = p90 time × cost against the best

Each provider's best-value machine (lowest p90 time × cost among its machines within 15% of its fastest of the same size) at 4 vCPU. Rust: RunsOn tested 9, Namespace 2, Ubicloud 2, the others 1; TypeScript: RunsOn tested 9, Namespace 2, Ubicloud 2, the others 1; Docker: RunsOn tested 8, Ubicloud 2, the others 1.

last run 13 Nov 2026 3–7 runs per build · 25 Oct–13 Nov 9 providers 94 runners

About this data

Harness: 58d5653
Reference size: 4 vCPU · 16 GB · variants 2–12 vCPU
Replicates: 3–7 runs per build per runner, 25 Oct–13 Nov · a failed run counts as a failure, not a missing run · 105 burst jobs
Jobs: 9923 of 10098 jobs finished across the Rust, TypeScript, Docker, hardware, burst and cache suites, 175 failed (Rust 3, TypeScript 76, Docker 3, Hardware 6, Burst 81, Cache 6)
Data: runs Rust, TypeScript, Docker, Hardware, Burst, Cache · runners.json, avrea.json, aws-codebuild.json, blacksmith.json, boat.json, github.json, namespace.json, runson.json, ubicloud.json, warpbuild.json · Methodology

Every runner builds the same three projects from a cold start: spotify-player (Rust), opencode (TypeScript) and PostHog image (Docker); values are medians across runs. The TypeScript typecheck runs 4 tasks at a time on every runner, so bigger machines gain less there.
Cost per build is variable cost: the whole job's duration times the provider's pay-as-you-go price per minute. RunsOn's starts from each job's own EC2 + EBS bill and is priced at one rate per mode, so both modes give the correct price: Spot (the default) puts every job at the spot rate, including any that fell back to on-demand because AWS had no spot capacity for the type (repriced at spot over the same billed time); On-demand puts every job at the on-demand rate. A figure never mixes the two. Plan fees and licenses are flat, listed on each provider's page.
The charts open on best value: time × cost per build, so a runner twice as fast at twice the price ties, and the × beside a bar is that product against the best one. Time and Cost rank by one figure.
Time is the whole job by default: the build steps plus the runner's start-up and teardown. Wait + job adds the wait for a runner, but that wait also holds how long GitHub takes to dispatch the job, which varies from run to run and no provider controls, so it is not the default. Cost is the job only.
They open at the reference size, 4 vCPU: each provider's best machine of that size; the line under a chart says how many machines each provider fielded. “Any” races every variant, including sizes only one provider ran (RunsOn 12 vCPU, Warpbuild 6 vCPU and Warpbuild 12 vCPU).
A bar ranks, shades and gets its × only from 3 finished runs; a failed build stays on the record with its reason.

How ranking works →

Beyond builds

The same runners, one question each.

Queue How long jobs wait for a runner in a burst, and how far build times move from run to run.
Bigger runners What doubling the vCPU did to each build's time and cost, on each provider's own runner line.
Head-to-heads Two runners of the same size, build for build, cost first.
CPU Fastest single-thread on x64: Namespace 8x32, 4,556 PassMark.
Cache 4 GB restore on x64, any size: RunsOn 6s (12 vCPU), then Blacksmith 9s (2 vCPU); GitHub's own service 31s (8 vCPU).4 GB restore on x64 at 2 vCPU: Blacksmith 9s, then Avrea 13s; GitHub's own service 1m 01s.4 GB restore on x64 at 4 vCPU: Avrea 10s, then Blacksmith 12s; GitHub's own service 44s.4 GB restore on x64 at 8 vCPU: RunsOn 11s, then Avrea 11s; GitHub's own service 31s.
Disk Small random reads (QD1) on x64: Namespace 8x16 leads with 107,149 per second, 60× the slowest disk (AWS CodeBuild, 1,794). That is a disk's limit, and a build often never reaches it.
EC2 storage NVMe, gp3 and tmpfs on the same instances, and what each does to the builds above.
All runners All 94 runners in one table, with the price / performance chart.
tmpfs The workspace in RAM instead of on a disk: what it does to the builds above, what it costs and where it does nothing.

Queue How long jobs wait for a runner in a burst, and how far build times move from run to run.
Bigger runners What doubling the vCPU did to each build's time and cost, on each provider's own runner line.
Head-to-heads Two runners of the same size, build for build, cost first.
CPU Fastest single-thread on arm64: Avrea 4 vCPU, 4,823 PassMark.
Cache 4 GB restore on arm64, any size: Avrea 11s (8 vCPU), then RunsOn 11s (8 vCPU); GitHub's own service 36s (4 vCPU).4 GB restore on arm64 at 2 vCPU: Avrea 12s, then RunsOn 19s; GitHub's own service 53s.4 GB restore on arm64 at 4 vCPU: Avrea 11s, then RunsOn 16s; GitHub's own service 36s.4 GB restore on arm64 at 8 vCPU: Avrea 11s, then RunsOn 11s.
Disk Small random reads (QD1) on arm64: Blacksmith 4 vCPU leads with 32,691 per second, 4.0× the slowest disk (GitHub, 8,119). That is a disk's limit, and a build often never reaches it.
EC2 storage NVMe, gp3 and tmpfs on the same instances, and what each does to the builds above.
All runners All 94 runners in one table, with the price / performance chart.
tmpfs The workspace in RAM instead of on a disk: what it does to the builds above, what it costs and where it does nothing.

Providers

Price per minute of a 4 vCPU runner on x64 (its shape beside each price); RunsOn at EC2 spot and on-demand.Price per minute of a 4 vCPU runner on arm64 (its shape beside each price); RunsOn at EC2 spot and on-demand.

Cost per build is variable cost only; licenses and plan fees are flat, on each provider's page. Estimate your bill