"We need faster builds, so let's get the faster chip," is a reasonable-sounding sentence that has cost a lot of teams a lot of money. If you're sizing Mac infrastructure for CI, you've probably had some version of this conversation, and if you've tried to settle it with a spec sheet, you already know how that goes: Apple publishes single-core numbers, reviewers publish multi-core numbers, and neither one tells you how long your build actually takes or what it costs you every month to run it.
The benchmark most people reach for measures something quite different from the work their team does all day.
We've just finished benchmarking the new M6 and M5 Pro Mac minis in our Las Vegas data center to help you decide which of our standard models suits which use case. The M6 wins on single-core by 9%, but it loses to the M5 Pro on the iOS build by 29% and a full-core render by 67%. Whether you're running a handful of build runners or a few hundred, that gap should inform which machine you buy.
What we tested
These are the four standard Mac mini models we offer, and each one suits a different kind of work, so the benchmarks below show which model we'd recommend for which use case. We ran all of them on macOS 27.0 (build 26A428):
The types of cores matter more than the total core count, and they explain most of what we found.
Core configurations
Both chips in this round use Apple's three-tier core design, with Super, Performance and Efficiency cores, but each chip combines them differently.
The M6 ships in a single twelve-core configuration: two Super cores, four Performance cores and six Efficiency cores. Half of that count is doing the light work, which works well for a machine handling a mixed desktop workload.
The M5 Pro comes in fifteen-core and eighteen-core variants, and neither one includes a single Efficiency core. The fifteen-core pairs five Super cores with ten Performance cores, and the eighteen-core pairs six with twelve. Every core in an M5 Pro is a Super or Performance core, so all of them contribute to a parallel job, while half of the M6's twelve cores are Efficiency cores built for lighter work.
If you're running a short job, you won't see much difference between the two chips, but on a long, heavily parallel job, such as a full build or a render, the M5 Pro's extra Performance cores account for most of the gap you'll see in the results below.
How we tested
We run four benchmarks, and each one answers a different question:
- XcodeBenchmark compiles a real iOS project with 76 CocoaPods dependencies, which makes it the closest thing we have to an actual CI job, so it's the result to look at if you're choosing a machine for Orka CI or iOS and macOS build runners
- Geekbench 6 CPU gives us single-core and multi-core numbers that compare cleanly across chip generations, which helps if you're weighing a new machine against one you already run, and the single-core score is a good guide to how responsive a machine feels for interactive work such as VDI
- Geekbench 6 Compute (Metal) measures GPU throughput using Apple's native graphics API, which matters if your workload leans on the GPU, such as ML inference, rendering or simulation
- Cinebench renders a 3D scene across every available core, and it's the most sustained, most parallel load in the set, so it's our closest stand-in for long-running parallel work such as large builds or several agent sandboxes running at once
Each test ran three times per machine with ten minutes of low CPU utilization in between, and we report the best of three. Run-to-run variance landed between 0.0% and 2.2%, so you can expect the same results if you run these tests yourself.
Two caveats before the numbers:
'Didn't XcodeBenchmark break on Xcode 27?' you may ask, and yes, it did. The project targets iOS 14.5 and Xcode 27 requires a deployment target of 15.0 or newer, so every build fails out of the box. We applied an open pull request that raises the deployment target and doesn't change any source code. Our times are consistent with each other and safe to compare within this table, but they're not comparable to the public XcodeBenchmark results.
Cinebench 2026 isn't comparable to earlier Cinebench releases either. Maxon reset the scoring baseline, so a 2026 score lands roughly three times higher than an R24 score for the same work. We're treating these scores as a new baseline, so you won't see them compared against our earlier Cinebench results.
Benchmark results
XcodeBenchmark
The M5Pro.XL compiles the same project 29% faster than the M6.S.
Geekbench 6
This is where a spec sheet will steer you wrong. The M6 takes single-core by roughly 9%, and if that were the only number in front of you, you'd buy the M6 and feel good about it. Multi-core tells you the opposite, with the M5Pro.XL ahead by 35%.
Cinebench 2026
On the most parallel workload in the set, that gap opens up to 67%.
How these compare to M4 and M2
Two results stand out when you put these against the machines already in our fleet.
The M5Pro.XL mini outperforms the M2 Ultra Studio on CPU work. It's 33.5% ahead on Geekbench multi-core and compiles our test project 15.4% faster, in a much smaller chassis and at a lower monthly cost. If your workload mostly builds and tests, the Studio is no longer the machine you need for the job.
Our entry-level M6.S lands within 2.3% of that same M2 Ultra on Geekbench multi-core. This machine now performs about as well as a former flagship on general compute, which shows how much four chip generations have improved.
Tier over tier, the jumps look like this:
GPU is the exception. The M2 Ultra Studio still beats the M5Pro.XL on Geekbench Compute by 38.9%, because Ultra chips carry far more GPU cores than any chip in the Mac mini line. If your workload is GPU-bound, whether that's ML inference, rendering or simulation, a Studio is still the right machine, and we'll publish M5 Ultra numbers once that hardware is available.
Why the gap widens
If you line up all three results, the pattern is consistent:
- Single-core Geekbench: M6 ahead by 9%
- XcodeBenchmark: M5 Pro ahead by 29%
- Cinebench: M5 Pro ahead by 67%
The more work you hand these machines, and the more of that work can be split across cores, the further ahead the M5 Pro pulls, and that tracks the core layouts almost exactly. Six Super and Performance cores paired with six Efficiency cores works well for a desktop handling background tasks, but eighteen Super and Performance cores do much better when you're compiling a large project or rendering a scene, because those efficiency cores add far less to the result than the core count suggests.
Sizing a CI fleet on single-core scores will point you at the wrong machine.
Does more memory help?
Not for these scores, no. The M6.S has 16GB and the M6.M has 32GB, and they're otherwise identical machines. Across all four benchmarks they finished close to each other, and on Cinebench they were 0.23% apart, which is well inside normal machine-to-machine variation.
We've now tested both machines under identical conditions, and the 32GB M6.M came out fractionally ahead of the 16GB M6.S on every measure. The largest of those margins was 1.5%, which is inside run-to-run noise.
The same pattern held for the M5 Pros, where the difference between the 48GB and 64GB machines tracks their core counts (15 against 18) rather than their memory.
You should still choose memory based on what your workload needs, such as running several VMs at once or handling a particularly large project. Memory also decides how many agent sandboxes fit on one machine, which we cover in the agent sandbox section below.
What about agent sandboxes?
We haven't benchmarked agent workloads on these machines, so this section is based on our CI results and on how agent workloads typically behave.
A CI job is short and ephemeral, it spins up, builds, runs its tests and gets torn down. Agent workloads are different, because they're long-running and hold state between steps, and teams often run several of them at once on the same machine.
Because of this, a different spec matters most. For CI, our benchmarks show that core count matters most and memory makes little difference. For agents, memory is what determines how many you can run on one machine at once, since each sandbox holds onto its share for as long as it's running, so memory matters more here than it did in our benchmarks.
What we found about core layouts still applies, since running several agents at once creates a sustained parallel load, which is the kind of work where the M5 Pro's all-performance-core design pulled furthest ahead, 67% on Cinebench against 29% on a single build.
A guide to our standard Mac mini models
M6.S | 12 cores, 16GB, 512GB
This is our entry tier, and it has the best cost per build in the lineup. It works well for VDI, and it's a good choice if you want the most throughput per dollar. For agent sandboxes, 16GB is the tightest fit in the lineup, so it suits a single agent more than several.
M6.M | 12 cores, 32GB, 1TB
The additional memory gives you room to run more VMs at once, although it won't raise your benchmark scores. This is the minimum we'd recommend for Orka CI, and it's a good fit for offshore development teams. The same extra memory makes it a reasonable starting point for running a couple of agent sandboxes side by side.
M5Pro.L | 15 cores, 48GB, 2TB
This is our recommendation for Orka CI, and it's well suited to agent sandboxes. It builds roughly 20% faster than the M6 tier.
M5Pro.XL | 18 cores, 64GB, 2TB
This is the fastest Mac mini we offer, for both CI and agent sandboxes. Choose this one if reducing build time is your main goal, or if you want the most agents running on one machine.
The M5 Ultra Studio configurations (S5.L and S5.XL) weren't part of this round, and we'll publish those numbers as soon as the hardware lands.
The results, summarized
- The M5Pro.XL is the fastest Mac mini we offer, compiling a real iOS project in 81.89 seconds
- The M6 takes single-core by 9% and then loses real workloads by 29% to 67%, so don't size a CI fleet on single-core scores
- Memory capacity didn't move any benchmark, so choose it based on what your workload needs
- Memory does determine agent sandbox density, so size it by how many agents you plan to run on each machine
- Full results for every configuration we've benchmarked live in our benchmark documentation
MacStadium is racking M6 and M5 Pro minis in our data centers now for enterprise use cases. If you need a fleet of 5 or more Macs, contact our sales team.