So Apple just did something weird. On August 25 they announced the new Mac mini, and instead of the M5 everyone was expecting, the base model jumped straight to M6. No standard M5 Mac mini ever existed. Apple didn’t announce one, didn’t cancel one, it just skipped it, and went from M4 to M6 in one move while the higher tier moved from M4 Pro to M5 Pro.
I’ve been going through the actual spec sheets and pricing pages for the last two days trying to figure out where the real value is, because on paper this looks like a simple two-chip choice and it’s really not. The base M6 starts at 899 dollars. The M5 Pro starts at 1,699 dollars. That’s an 800 dollar gap, and depending on what you actually do with the machine, that gap either matters enormously or barely at all.
This isn’t a spec-sheet reprint. I’m going to walk through what changed, what the early benchmark numbers actually say (a lot of them are still estimates, I’ll flag which), and where I’d put my own money if I were buying today, whether I own nothing, an M1, or an M4 Pro sitting on my desk right now.
The lineup, in plain terms
The M6 Mac mini has a 12-core CPU (2 super cores, 4 performance cores, 6 efficiency cores) and a 12-core GPU with Neural Accelerators built into every core. Memory tops out at 32GB, bandwidth tops out at 170GB/s, and you get three Thunderbolt 4 ports. Base config is 16GB RAM and 256GB storage for 899 dollars.
The M5 Pro steps up hard. Up to an 18-core CPU, up to a 20-core GPU, up to 64GB of unified memory at 307GB/s bandwidth, three Thunderbolt 5 ports, and genlock support over USB-C for people syncing cameras and displays. It starts at 1,699 dollars for a 15-core CPU / 16-core GPU config with 24GB RAM.
Here’s where it gets weird if you’re actually configuring one. RAM and storage are locked in at checkout, no upgrading later. On the M6, going from 16GB to 32GB costs 400 dollars. On the M5 Pro, going from 24GB to 64GB costs 1,000 dollars. Daring Fireball put together the full pricing tables the same day Apple announced this, and reading through them, the jump from the entry M5 Pro (1,700 dollars, 15-core, 24GB) to the maxed 18-core/20-core version with 64GB pushes past 2,900 dollars before you’ve added storage. That’s Mac Studio territory, and worth knowing before you get lured in by “just a small bump.”
What the benchmarks actually show
None of this has shipped yet. Deliveries start September 22, and no independent Geekbench results exist for M6 or M5 Pro as of today. Everything below is either Apple’s own claim or a calculated estimate, and I want to be upfront about which is which instead of presenting projections as facts.
For context on where we’re coming from: the 2024 Mac mini with M4 Pro averages 3,311 single-core and 24,711 multi-core on Geekbench 6, based on 125 uploaded results. The M2 Pro model before that sat at 2,403 single-core and 16,243 multi-core. That’s roughly the generational jump we’re talking about, and it’s been fairly consistent release over release.
AppleInsider ran the math using Apple’s own claimed improvements. Taking the M4 Mac mini’s Geekbench 7 single-core score of 3,277 and applying Apple’s stated 40 percent CPU gain gets you to about 4,590 for the M6. Multi-core, using Apple’s 1.2x multithreaded claim, goes from 15,359 to roughly 18,431. On the graphics side, Apple says the M6 GPU is twice as fast as the M4’s, which would take a Metal score from 54,176 up past 100,000, putting it in the same range as the outgoing M4 Pro. I’d treat these as directionally right, not as numbers to plan a purchase around.
Real testing will settle this in a few weeks. Until then, the honest answer is: probably a meaningful CPU bump, a big GPU bump, and the fine details are still anyone’s guess.
Something that confused me for a good ten minutes, so I’ll save you the trouble: Geekbench scores from different major versions aren’t directly comparable. A Geekbench 6 multi-core score of 24,711 and a Geekbench 7 score for the same chip aren’t on the same scale, Geekbench 7 recalibrated its baseline against an AMD Ryzen 7 7700 at 2,500 points. So when you see the M6 estimate of 4,590 single-core sitting next to an older M4 Pro’s 3,311, don’t assume that gap is real, one number is Geekbench 6, the other is a Geekbench 7 projection. Most tech sites don’t clarify this and it makes side-by-side charts look more dramatic than they are.
Editing, video, and the stuff creators actually care about
If you cut video or do 3D work, the M5 Pro is the one built for you, and Apple isn’t shy about it. Popular Science pulled the DaVinci Resolve claim directly: up to 3x faster Magic Mask performance versus the M4 Max, which is a strange comparison since M4 Max was never in a Mac mini, but it tells you where Apple thinks this chip sits. Blender users get a quoted 1.4x ray tracing improvement over M4 Pro.
Genlock is the feature nobody was asking for and photographers are going to love anyway. It lets you sync a display and a camera precisely over USB-C, including the iPhone 17 Pro, which matters if you’re doing multi-camera shoots or timelapses where frame drift ruins the take. It’s M5 Pro only. The M6 doesn’t get it.
Thunderbolt 5 on the M5 Pro also opens up clustering multiple Mac minis together for heavier workloads, something a few studios were already doing with M4 Pro units over Thunderbolt 4, just slower. Storage got faster too across both chips, roughly double the previous generation according to Apple, which shows up as shorter export and project-load times if you’re working off internal SSD rather than external drives.
If your editing work is lighter, quick color passes, short-form video, basic Lightroom batches, the M6’s 2x graphics claim over M4 probably covers you fine. I wouldn’t pay the 800 dollar M5 Pro premium just for genlock unless you know you need it.
Now the part I actually got into this for: local LLMs
This is the part I actually got curious enough to dig into properly. It also needed me to slow down and separate hype from what’s actually measurable.
Apple is marketing the M6 hard around on-device AI. Dual 16-core Neural Engine (first time on any Mac mini), Neural Accelerators in every GPU core, and a claimed 4.8x faster time-to-first-token in LM Studio versus M4, 13.5x versus the original M1. The M5 Pro claims 4x versus M4 Pro and 8.5x versus M2 Pro on the same measure. These are Apple’s own preproduction numbers, tested on a 32GB unit, and Apple hasn’t published the model or quantization used, so treat the multiplier as marketing copy until someone reruns it independently.
It took me longer than I’d like to admit to properly understand this, but it matters more than the flashy multiplier does.Local LLM generation speed is bandwidth-bound, not compute-bound. The chip has to physically stream model weights out of memory for every token it produces. So the raw core count matters less than how fast that pipe is. M6 pushes bandwidth to 170GB/s, up from 153GB/s on M5-class silicon and 120GB/s on M4. That’s a real improvement but nowhere close to 4x, no matter what the marketing chart implies.
MacStories ran a rough calculation worth quoting because it’s the clearest real-world translation I’ve seen. A mixture-of-experts model like Qwen 3.5–35B-A3B runs at around 17 tokens per second on a base M4 Mac mini with 16GB. If the M6’s gains scale roughly linearly with Apple’s claims, that same model could realistically hit over 60 tokens per second. That’s the difference between a chatbot that feels sluggish and one that keeps pace with how fast you read.
But memory ceiling is still the real gatekeeper, more than the chip name on the box. The M6 tops out at 32GB. The M5 Pro goes to 64GB with the better 307GB/s bandwidth. If you want to run large 70B-class models, the M6 simply can’t, no matter how fast its Neural Engine is. 101 AI Tools, which runs a site tracking Apple Silicon for local AI use, put it well in their M6 vs M5 Pro breakdown: buy memory for the models you actually want to run, and treat Pro as the tier for bandwidth and headroom, not as an automatic upgrade.
I’ll admit my own bias here. I mess around with local models mostly for small coding assistants and document search, nothing that needs 70B parameters, so a 32GB M6 configured with the 400 dollar RAM bump is plenty for me. If your plan involves running something like a 70B chat model at home, you’re not shopping M6 at all, you’re shopping M5 Pro at 64GB minimum, and honestly you should also be looking at the M5 Max Mac Studio, since that’s a different conversation entirely.
There’s a dataset floating around called LLMCheck that tries to put real numbers to all this instead of just marketing multipliers, pulling 227 tokens-per-second figures across 58 models and 16 different Apple Silicon chips, from Ollama, LM Studio, and MLX runs. Some of those are vendor numbers, some are community submitted, some are estimated from the bandwidth math, and each row is labeled which is which, which I appreciated since so much of this space just blends claim and measurement together.
Their rough conclusion tracks with what I said above: small models under 9B run fine on basically anything, even an old M1 with 16GB gets 40 to 80 tokens per second. Mid-size 14B to 35B models want something like an M4 Pro with 24GB as the practical floor. Past 70B, you’re not looking at a Mac mini at all anymore, you want the memory ceiling only Mac Studio’s M5 Max or M5 Ultra chips offer.
Going further back: M1 and M2 owners
There’s no M3 Mac mini, Apple skipped that one for the desktop line too, so this section covers M1 and M2 only. A lot of people still running the 2020 or 2023 model keep asking me if it’s finally time. Let’s put numbers next to the years instead of guessing.
The Late 2020 M1 Mac mini scores about 8,900 single-core on Geekbench 6. The 2023 M2 Pro model jumped to 2,403 (Geekbench 6 uses a different baseline scale than the M1-era chart, which is part of why Apple’s own comparisons jump between different multipliers depending on which generation they’re measuring against, and it trips people up constantly). What matters practically is this: anyone still on M1 is looking at roughly four to five years of chip generations stacked up against them by the time the M6 ships. That’s not a subtle difference.
Apps launch instantly instead of bouncing in the dock, multitasking with quite a few Chrome tabs and an editing app open stops causing fan noise, and Apple Intelligence features that need on-device processing actually run at a usable speed instead of falling back to cloud processing or just not working. I still keep old M1 mini around for backups, and honestly the fan kicking in the moment I open two heavy apps together is the real tell. It’s not slow exactly. It’s just old, and it shows the moment you push it even slightly.
M2 owners are in a slightly better spot but the case for upgrading is still strong, mostly because of the AI angle more than raw CPU speed. The M2 generation predates Apple’s heavier push into on-device models entirely, and memory bandwidth on those chips sits well below even the base M6’s 170GB/s.
A quick real-world example
A friend of mine runs a small video editing setup, mostly wedding highlight reels, and she’s been on an M2 Pro Mac mini since early 2023. She wasn’t planning to upgrade this year at all. Then Adobe pushed a Premiere update in July that added AI-assisted rough cuts, and her exports started taking noticeably longer, like almost double, because the M2 Pro doesn’t have the newer Neural Accelerator hardware the feature seems to lean on. She preordered the M5 Pro the day it opened, the 64GB configuration, and is already annoyed the shipping estimate slipped to early October. I don’t think everyone needs to follow her lead, her job depends on that one specific feature. But her M2 Pro itself was never broken. It just got left behind by a software update, not a hardware failure.
Should you actually buy it, or upgrade?
Break this down by what you’re currently sitting on, because the right answer changes completely depending on where you’re starting.
If you’ve got nothing right now, just get the M6. For most people, an 899 dollar computer with 40 percent more CPU headroom and double the graphics of the outgoing model covers browsing, office work, photo editing, and light coding without breaking a sweat. Bump the RAM to 24GB or 32GB if you can stretch the budget, the base 16GB config is fine today but you’ll feel it cramped in two years, especially if Apple Intelligence features keep expanding.
If you’re on an M1 or M2 Mac mini, upgrade, easily. The generational jump here is large enough that you’ll notice it in everyday use, not just benchmarks. An M1 Mac mini scores around 8,769 single-core on Geekbench 6. Even the conservative M6 estimates put it well past double that.
M4 or M4 Pro from 2024 is the harder call, and I don’t think there’s one clean answer for everybody. If your machine still does what you need, skip this cycle. A 40 percent CPU claim and a doubled GPU is real, but it’s not the kind of leap that makes a two-year-old machine feel broken. The one exception is if local AI is specifically why you’re upgrading. Then the memory bandwidth change is worth chasing, and the M6’s dual Neural Engine is a properly new piece of silicon, not a rebrand.
Need 64GB, Thunderbolt 5, or genlock for professional work? Then it’s M5 Pro, no real debate. The premium buys you actual capability the M6 physically cannot deliver, not just a faster version of the same thing.
I lean toward the M6 being the smarter buy for most people reading this, maybe 75/25 over the M5 Pro. The M5 Pro’s extra power is real, but it’s aimed at a narrower set of jobs, video production, 3D work, running big local models, professional AV syncing, and most people don’t have those jobs. Where I’d push back on Apple a little: the 899 dollar starting price is a 100 dollar jump from the outgoing generation’s base price, and 256GB of storage at that price still feels stingy in 2026. One 9to5Mac commenter called it customer abuse, and while that’s a bit dramatic, they’re not wrong that 256GB fills up fast once you’re running local models and editing photos on the same machine.
One thing nobody can answer yet, including Apple: whether any of these performance claims hold up once real Geekbench and LM Studio numbers start showing up after the September 22 ship date. If you’re not in a rush, waiting three or four weeks for independent testing costs you nothing and tells you a lot more than a press release chart ever will.