Avrea vs Blacksmith

Avrea and Blacksmith both run your GitHub Actions jobs and charge the same entry-level x64 minute rate. Avrea has the highest measured ARM CPU scores in the 4-vCPU benchmarks below. Its build caches reduce repeated work, while Avrea Quality proposes and tests CI fixes.

Try Avrea

Features at a glance

Both keep GitHub Actions. Supported Actions cache requests are routed automatically on both; Docker caching and AI tools work differently.

Avrea vs Blacksmith: feature comparisonScroll sideways to compare both providers.
FeatureAvreaBlacksmith
Operating systemsLinux, Windows and macOSLinux, macOS and Windows in public beta
Linux x64, 2 vCPU / 8 GB$0.004/min$0.004/min
Linux ARM, 2 vCPU$0.012/min, 8 GB$0.0025/min, 6 GB
Actions cacheAutomatic integration with actions/cacheAutomatic integration with actions/cache
Docker cachingPersistent disks and a BuildKit cache endpoint for Docker cachingPersistent NVMe builder state and cache mounts
Build-tool cachesReuse outputs with integrations including Turborepo, Gradle, Go and BazelReuse Bazel build outputs across jobs and branches
DebuggingSearchable logs, resource metrics and SSHLogs, resource metrics and SSH
CI analyticsWorkflow and job failure/flake trends, plus cache diagnosticsTest analytics from JUnit reports and supported test logs
Static IPsEnterprise feature; $100/IP/monthPaid add-on; $100/IP/month
AI-generated pull requestsAvrea Quality proposes CI fixes, tests them and opens PRs[code]smith makes code changes in PRs, runs GitHub Actions checks and iterates on failures

Rates exclude included usage and paid extras.

Runner pricing

Linux, 2 vCPUAvreaBlacksmith
x64$0.004/min · 8 GB$0.004/min · 8 GB
ARM64$0.012/min · 8 GB$0.0025/min · 6 GB

Avrea’s Linux ARM runners use Apple M5 Max hardware. They cost more per minute, but Avrea has the fastest measured ARM CPU performance in the 4-vCPU PassMark tests below. Avrea also includes more memory at the 2-vCPU size above. Avrea pricing, Blacksmith pricing, Blacksmith specifications.

Both advertise 3,000 free monthly minutes, with larger runners using more of the allowance. Blacksmith counts those minutes in x64 2-vCPU equivalents, with different conversion factors for ARM, Windows and macOS. Avrea plan details, Blacksmith allowance rules.

Our GitHub Actions pricing guide explains how runner time, included minutes and storage contribute to the bill.

Compare the bill for a completed job

At the listed 2-vCPU x64 rate, a ten-minute job costs $0.04 on either provider before allowances or extras. If one finishes that same job in six minutes, its compute cost is $0.024. With equal minute rates, elapsed time and cache charges decide the difference.

CPU and cache benchmarks

These results come from RunsOn’s public GitHub Actions benchmarks. RunsOn sells runners too, and publishes its test code and methodology. The tables below show the 4-vCPU results for both providers.

CPU

Medians from five runs per runner, 14 October 2026.

4-vCPU runnerAvreaBlacksmith
x64 · PassMark single-thread4,2434,230
x64 · PassMark CPU Mark, all cores11,17414,759
ARM64 · PassMark single-thread4,8601,346
ARM64 · PassMark CPU Mark, all cores8,3883,778

Avrea has the highest ARM scores across all tested providers at 4 vCPUs. Namespace’s Apple-silicon runner is close, within 3% on both measures. Blacksmith has the higher x64 multicore score, while the x64 single-thread scores are close. These are CPU tests, so the scores don’t translate directly into build times. RunsOn CPU results.

RunsOn lists 16 GB for all four runners, while Blacksmith’s catalog lists 12 GB for its 4-vCPU ARM runner.

Actions cache

This test saves a 4 GB random file, deletes it, then restores it immediately on the same runner.

4-vCPU runnerAvreaBlacksmith
x64 · restore / save14 s / 20 s15 s / 7 s
ARM64 · restore / save11 s / 15 s34 s / 21 s

On ARM, Avrea restores the file in 11 seconds; Blacksmith takes 34. On x64, restore times are close and Blacksmith saves faster. This test measures Actions cache transfers, not Docker caching. RunsOn cache results.

The cache table uses the 14 October 2026 results, with five runs per runner shown here. Our guide to runner benchmarks explains what to measure beyond CPU scores.

Data: RunsOn, CC BY 4.0. We selected the 4-vCPU rows and shortened the metric labels.

Caching

Both providers route supported actions/cache usage to their own cache automatically. Blacksmith documents a free 25 GB repository allowance. Avrea includes 25 GB per repository, shared across Actions, LFS, build and package caches, with additional cache storage available to purchase. Blacksmith Actions cache, Avrea cache storage.

Avrea caches both downloaded packages and build outputs. Its Turborepo integration reuses valid task outputs, so a cache hit can skip compilation or tests rather than just speed up a download. The remote-cache settings are already configured on Avrea runners. Package proxies serve repeat downloads from registries such as npm and PyPI. Avrea Turborepo cache, package caching.

Avrea also integrates with Gradle, Maven, Nx, Nix, ccache, sccache, Go, Bazel, Tuist and Xcode. These use each tool’s cache settings: enable build caching in Gradle, for example, or install and use sccache for Rust. Build-cache integrations, Gradle setup, sccache setup.

Blacksmith also has build-tool caching. Its Bazel integration reuses outputs across jobs and branches within a repository. These caches can skip compilation or test work entirely, which the Actions cache transfer benchmark above doesn't measure. Blacksmith Bazel cache.

Avrea’s persistent Docker cache preserves images, layers and builder state between jobs, reducing the work a repeat build needs to do. You can also export layers to Avrea through BuildKit’s type=gha backend. The Docker guide covers both options, including the persistent-disk settings and the url_v2 cache endpoint. Avrea Docker setup.

Blacksmith’s Docker integration keeps builder state on persistent NVMe storage, including layers and cache-mount data. Docker layer caching and sticky disks are listed at $0.50/GB/month. Blacksmith Docker setup, pricing.

For a walkthrough of cache keys, restore behavior and invalidation, read our GitHub Actions caching guide.

Package caches and build outputs

A dependency download, a compiler invocation and a Docker layer rebuild are different costs. Restoring a package directory can save network time while leaving compilation untouched. A build-tool cache can reuse an output and skip that compilation. Persistent Docker builder state can also preserve data that a normal Actions cache test never touches.

In a TypeScript monorepo, package caching reduces dependency downloads, while Turborepo or Nx can reuse outputs from unchanged tasks. Persistent Docker builder state also retains cache-mount data between builds.

Debugging and AI-assisted fixes

Avrea includes searchable logs, CPU and memory metrics, and SSH access to running jobs. Blacksmith also has step-level resource metrics and SSH access. Avrea logs and metrics, Avrea included features, Blacksmith metrics, Blacksmith SSH.

Avrea’s workflow and job analytics track failure rates, flake rates and duration trends. From a slow or failed run, you can inspect the logs and cache hits, misses, transferred bytes and errors. Blacksmith’s test analytics collects JUnit reports and parses supported test output from logs to show individual test results. Avrea analytics, Avrea cache diagnostics, Blacksmith test analytics.

CI fixes and coding agents

Both Avrea Quality and Blacksmith's [code]smith use CI results to work on code changes. With Avrea Quality, the work starts with a failing test or a problem identified in the pipeline. It uses the logs, diff and test history to propose a change, runs your full test suite and opens a PR if it passes.

With Blacksmith's agent, you can hand over a broader coding task. It pushes changes to a pull request, waits for GitHub Actions checks, reads failed-job logs and makes further changes. When the jobs run on Blacksmith, it can also inspect their resource usage. Blacksmith's runner sizing recommendations can produce PRs too, so automated pipeline improvements are part of both products.

Avrea Quality starts from CI data; [code]smith also accepts broader coding tasks. The Avrea CLI and Blacksmith CLI give coding agents access to job information and logs.

Avrea also has sandboxes in beta. Blacksmith's Testboxes, in early beta, let you run commands against local changes inside a GitHub Actions job.

Network access and dependency records

Both Avrea and Blacksmith offer static IPs for IP allowlisting, at a published price of $100 per IP per month. Avrea lists static IPs under its Enterprise features; Blacksmith sells them as an add-on enabled through support. Avrea pricing, Blacksmith pricing, Blacksmith setup.

Avrea's Enterprise plan includes an egress firewall. You can allow or block destinations and see which connections each job made or attempted. The rules are enforced outside the runner VM. Traffic is allowed by default until restrictions are configured. Required GitHub services and Azure Blob Storage remain reachable even with a deny-by-default policy. Firewall behavior and exceptions, plan availability.

Avrea also generates a software bill of materials (SBOM) for each repository, tied to a commit. It lists dependency versions and licences, with CycloneDX and SPDX exports. Avrea builds this from manifests and lockfiles, so it covers the repository's declared production dependencies. Avrea repository SBOMs.

What changed for GitHits

After moving from Blacksmith to Avrea, GitHits reported 27% faster median test feedback and a 34% improvement at p90, while its test suite grew 22%. The share of Docker builds finishing in one to two and a half minutes rose from 45% to 73%.

In a separate comparison of an identical job at the same vCPU size, Avrea removed about 15 seconds of runner platform overhead per job. The compute steps took the same time.

The feedback and Docker figures compare a month of successful runs before and after the migration. GitHits also split its pipeline into parallel stages during the move, so those gains reflect the runner change and the pipeline work together. Read the GitHits case study.

Choosing between Avrea and Blacksmith

At the same entry-level x64 minute rate, Avrea gives you build-output caching, diagnostics for slow and failing jobs, and tested fixes through Avrea Quality. For ARM workloads, its CPU results above show the performance available at the higher minute rate.

At GitHits, less runner overhead and more effective parallelism shortened the wait for feedback even when an individual job’s compute steps took the same time.

Try one job on Avrea

Connect your repository using the setup guide, then change the runner label on one job. For the standard 2-vCPU / 8-GB Linux x64 shape, the label is avrea-ubuntu-latest. Choose a matching explicit size if the Blacksmith job currently uses a larger machine.

Run the same commits on both providers. Include a cold cache, a repeat of the same commit and a typical source change that invalidates part of the cache. Record queue time, total job duration, the slowest steps and compute plus storage charges. Keep retries and failed jobs in the cost calculation.

Try Avrea · Runner sizes and pricing

Written by Avrea. Prices, documentation and benchmark results checked 14 October 2026.