The Mac Studio M5 Max is out, and what the first benchmarks say about gaming and AI

Apple's Mac Studio with M5 Max went on sale on September 22, 2026. It starts at $2,499 with an 18-core CPU, a 32-core GPU, 36GB of memory and a 512GB SSD, which is $500 more than the M4 Max model cost at launch. The first reviews agree on one slightly odd headline: the biggest gain is the SSD, not the chip. The gaming and local AI numbers are good too, but nearly every one of them comes from a 40-core GPU, 128GB machine that costs $6,899.

Most of the record comes from one review, published on September 21 by Roman Loyola under both the Macworld and PCWorld mastheads. His review unit had an 18-core CPU (6 super cores, 12 performance cores), a 40-core GPU, a 16-core Neural Engine, 128GB at 614GB/s and a 4TB SSD. Geekbench 7 had only just come out, so his comparisons with older Studios use Geekbench 6 where he had no newer results. Macworld gives percentages, not raw scores, for most tests. Raw scores come from tbreak.com, which tested the same chip with an 8TB SSD on macOS 27.0. The table puts them side by side, and every row names its source. Figures from different reviewers are different tests, so read across a row only where the source is the same.

The SSD is the story because it beat Apple's own number. Apple advertised storage up to 2x faster, and MacRumors' roundup of the Blackmagic results puts reads at about 2.6x and writes at about 2.1x the M4 Max Studio. A vendor claim that turns out conservative is rare enough to lead a review.

Gaming. tbreak's 193.67 fps in Cyberpunk 2077 is more than double the Mac mini M6 (82.2 fps) and the Mac mini M4 Pro (83.5 fps) in the same test. It ran at 1080p with High textures and MetalFX Quality upscaling, without frame generation or ray tracing. That is a lighter load than the Ultra presets PC cards get measured at, so it does not line up against a Windows chart. Macworld found gains of 30% or more in Warhammer III and Borderlands 3, and nothing in Civilization VI.

Local AI. Two things are going on, and they point in opposite directions. In Geekbench AI, which runs on the Neural Engine, the $899 Mac mini M6 beat the Studio at every precision in tbreak's tests (quantized: 85,175 against 67,954). Notebookcheck's early leaked scores (page updated October 5, 2026) showed only about 15% (half precision) and 10% (quantized) over the M4 Max. Its author argues the benchmark may not yet use the neural accelerators built into each M5 GPU core. Real LLM runtimes go through the GPU and memory bandwidth, and there the Studio is well ahead. In tbreak's MLX tests, Qwen3 30B-A3B prompt processing ran at 5,402 tok/s against 1,752 on the M6 mini, and generation at 130.6 against 62.0. Llama 3.3 70B at 4-bit peaked at 41.2GB of memory and generated 13.1 tok/s. Under Ollama, GPT-OSS 120B ran at 75.65 tok/s. The RTX 5090 could not load it, because 32GB of VRAM is too small, but on Qwen3 30B A3B it ran at 296.62 tok/s, which is 2.6 times the Studio's 113.74.

Hardware Corner published Reddit-sourced numbers for the same chip in a 128GB MacBook Pro. gpt-oss-120B at 8-bit with a 4K context generated 87.9 tok/s, against 221.3 on an RTX Pro 6000. The site's verdict was blunt: "The M5 Max is not replacing high-end GPUs." Apple says the chip processes LM Studio prompts up to 10.7x faster than the M1 Max and 3.9x faster than the M4 Max. Those multipliers come from Apple's own testing, and no reviewer in the record has reproduced them. Checking them would take someone running the same LM Studio model on all three machines. The site's LLM inference table collects the cited runs.

Against the MacBook Pro. None of these sources measures sustained clocks on the Studio. Notebookcheck's review of the MacBook Pro 14 M5 Max (March 11, 2026) ran under a headline about "inconsistent performance and throttling issues". The bigger chassis is the obvious fix, but nobody has published a measurement showing it.

Ports. Four Thunderbolt 5 ports (80Gbps, up to 120Gbps with a compatible cable and device), two USB-C, two USB-A, HDMI 2.1, 10Gb Ethernet, an SDXC slot and a headphone jack. Wireless is Wi-Fi 7 and Bluetooth 6. Under macOS 27 it drives up to five external displays, or three at higher resolution (Macworld).

Apple announced the line on August 25 and opened pre-orders the same day. It put the higher starting prices down to an ongoing memory chip shortage (MacRumors). The increases started before this launch: the M3 Ultra Studio had already gone from $3,999 to $5,299 in June 2026, and the M5 Ultra now starts at $5,499 with 96GB (PCWorld). The shortage also shows up in the upgrade sheet. Going past 36GB of memory adds at least $600 (Macworld), and memory is exactly the spec a local AI buyer can't skimp on.

The base model has a 32-core GPU, and no review in the record tested one. Every GPU, gaming and LLM figure above comes from a 40-core chip with 128GB. The memory figure is the one that changes the buying decision.

From an M4 Max: the gains are a 22% higher Geekbench 7 multi-core score, 33% in Cinebench 2024 GPU and about 2.6x on SSD reads. Loyola argues the storage alone justifies the move. That holds if your work is bound by file I/O; for gaming alone it is a harder sell.

From an M3 Ultra: stay put if you do Metal compute or heavily multi-threaded work. The older Ultra still leads Geekbench 7 multi-core by 9% and wins the Metal scores with twice the GPU cores. The M5 Max's advantages are the SSD and a 10% lead in Cinebench 2024 GPU.

Price: TechRadar reported a $200 cut at B&H Photo, taking the base model to roughly $2,299, which is the same as Apple's education price. Walmart prices aren't in the record. For Amazon, check the listing yourself. Against the cheaper desktop, the Mac mini M5 Pro comparison shows what the extra money buys.

Nobody in these sources has published a raw Geekbench 7 single-core score for the Studio, raw Cinebench scores, HandBrake or iMovie export times, or Blackmagic MB/s figures. Macworld's results for those exist only as percentages and chart images. These sources also have no creative-app results against the M3 Ultra, no sustained-clock data and no results for the 32-core base GPU. The Geekbench AI single-precision result that came in below the M4 Max hasn't been explained. Notebookcheck's theory about the GPU accelerators is the author's own reading, not a confirmed cause.

Macworld said the 512GB M5 Ultra option was due in October 2026, and MacRumors lists it at $22,299. Look for independent LLM runs on it, the first real check of Apple's claim of 4.3x the M3 Ultra's peak AI compute. Watch for a Geekbench AI update that uses the M5's GPU neural accelerators, which would revise the Neural Engine rows above (no date has been announced). Macworld also passes on an unconfirmed rumor that M7 chips will focus on local AI. There is no date for that either, and it is no reason to hold off on buying a Mac you need now.