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

Apple released the Mac Studio with M5 Ultra on September 22, 2026, at $5,499 for a 30-core CPU, 64-core GPU, 96GB and 1TB. The published benchmarks agree on the shape of it: a large jump in local AI over the M3 Ultra, a solid jump in CPU, and gaming that still trails dedicated-GPU PCs in most titles. Almost every number below comes from a 36-core, 256GB machine, which cost $11,299 to $12,299 in the reviews that priced it, so the entry model is not what was measured.

The top chip has a 36-core CPU (12 super, 24 performance), up to an 80-core GPU, up to 512GB of unified memory and 1.2TB/s of memory bandwidth, with UltraFusion at 4.4TB/s, as September launch coverage reports it. Only 96GB and 256GB were available at launch, Engadget noted.

Apple claims up to 1.25x single-threaded and 1.3x multithreaded CPU performance over the M3 Ultra (MacRumors, Sept. 15). It also claims 1.8x graphics over the M3 Ultra, up to 4.3x peak AI and up to 9.8x prompt processing against the M1 Ultra (iPhoneSoft, Sept. 21). Launch coverage also carries 4x for prompt processing, 1.7x for Redshift, and 5.3x and 3x for DaVinci Resolve Magic Mask, each set against the M3 Ultra or M1 Ultra.

These are vendor numbers, and the 9.8x gets its size from the age of its baseline. Against the M3 Ultra, MacStories measured prompt processing at roughly 2.5x on average and BestHub measured 3.6x at eight concurrent streams. The CPU claim holds up better: Macworld has Geekbench 7 multi-core about 34% ahead.

Each row below is one publication's own run of both machines. Don't read across publications.

Across the wider line, MacRumors' Geekbench 7 multi-core ladder (Sept. 15) runs M5 Ultra 52,516, M3 Ultra 39,059, M5 Max 35,119, M5 Pro 33,593, M1 Ultra 23,328, M1 8,582. In Geekbench 6 multi-core, Tom's Guide has 37,777 against 26,966 for the M4 Max, and its Mac mini M6 review has that machine at 21,405 (Sept. 21). The Verge (Sept. 21) recorded Geekbench 6 at 4,310 single-core, 37,870 multi-core and 233,865 in GPU Metal. These sources give no M2 Ultra figures and no M1 Ultra single-core score.

The spread between sources is partly benchmark version. TechRadar's roundup puts the Ultra about 26% ahead of the M5 Max in Geekbench multi-core. Macworld has about 47% in Geekbench 7 and about 25% in Geekbench 6, and warns the two versions are not directly comparable. A buyer quoted "26%" should ask which one.

Apple never shipped an M4 Ultra, so this generation jumps straight from the M3 Ultra, MacRumors notes. The first Geekbench 7 result surfaced a week before launch under the identifier Mac17,15. The price moved with the silicon: Engadget has the base Ultra at $5,500, up from $4,000, and a configuration comparable to its $11,299 review unit at $7,499 last generation. Memory bandwidth rose from 819GB/s to 1.2TB/s, per MacStories, and that number explains most of what follows.

Local AI. Tom's Hardware (Sept. 2026) found token throughput nearly 4x Nvidia's DGX Spark and roughly 2x the M4 Max, and credits bandwidth: 1.2TB/s against the Spark's 273GB/s. The RTX 5090 still wins while a model fits in its 32GB. MacStories had it reading a 6K-token prompt at about 3,000 tok/s against about 1,700, and generating at 59 tok/s against 48, with the M3 Ultra at 31. Past 32GB the card offloads to system RAM over PCIe, which Federico Viticci calls "no way to live." That capacity is the Mac's argument, not raw speed.

Two limits. BestHub's eight-stream aggregate decode is flat against the M3 Ultra, so the gain under concurrent load is prompt processing, cutting its pp1024/tg512 run from 32.971 s to 12.176 s. And 256GB has a ceiling: MacStories fit a 5-bit Flash-Next quant in RAM, while 6-bit and 8-bit needed SSD offload. Check how much memory a model gets before paying Tom's Hardware's quoted $4,000 to go from 96GB to 256GB. On images, MacStories logged Qwen-Image-2.1 at 180 seconds an image with 78GB peak RAM, which is one-third of an image a minute.

Gaming. Tom's Guide (Sept. 2026) ran three titles against an Alienware Area-51: Assassin's Creed Shadows 51 fps vs 140, Borderlands 3 73 vs 119, and Cyberpunk 2077 77 vs 56. Cyberpunk is the only win. Engadget (Sept. 2026) saw Baldur's Gate III average over 80 fps at 4K Ultra. BestHub called 37 fps with full ray tracing "playable but not smooth," and TechRadar's survey of 10 reviews named gaming the "biggest weakness."

The synthetic graphics results are closer. iPhoneSoft (Sept. 21) reports Blender at 15,232 samples/min against 14,966 for a Core Ultra 9 285K and RTX 5080 PC, and Steel Nomad at 8,073 against 8,221. That is a near tie with a card that launched at $999.

M5 Max or M5 Ultra. Stay on the M5 Max unless the work scales with cores or bandwidth. Engadget has single-core matching the Max and Geekbench AI only about 10% ahead. Macworld has Cinebench 2026 about 72% higher on CPU and about 49% higher on GPU. M4 Max owners get the 2x token figure and little reason to move otherwise. Non-professionals get the $899 Mac mini M6, on Tom's Guide's recommendation. Noise is not a factor: Tom's Guide reports no heat or fan trouble under sustained load, and Viticci could hear the fan only with an ear on the case.

The entry model. Every review here that names its configuration used a 36-core, 256GB machine. None reports the 30-core, 96GB box most people will price up first.

The 512GB build. AI Weekly said the option was due in late October with no price announced. Macworld estimated $5,000 or more above $12,299, and launch briefs put it near $22,299. Those figures are thousands of dollars apart, and these sources contain no 512GB benchmark. Some briefs also attach $6,799 to a 256GB machine; MacRumors gives it only as the 36-core starting price, and Engadget's 256GB unit cost $11,299.

Geekbench 7 GPU. MacRumors found no result on Sept. 15. Later September briefs repeat Metal scores of 352,444 to 360,019 against 255,009 to 263,242 for the M3 Ultra, or 33% to 41%, without one review we can cite for them. Engadget describes GPU scores as roughly 20% to 30% higher. Separately, The Verge found rendering gains up to 63%, as reported by iPhoneSoft.

Raw tokens from Tom's Hardware. AI Weekly's summary gives the 4x and 2x multiples and no tok/s figures, so those ratios have no absolute numbers here.

Apple's clustering claim. Apple says four M5 Ultra units run AI workloads 3x faster than one. Engadget could not verify it, and verifying it takes $22,000 of base-model Macs at minimum.

BestHub's reference scores. The write-up says its 3DMark scores exceed an RTX 5080 at 8,800 and an RTX 5070 Ti at 27,915. Its own results are 8,052 and 26,105. Treat that as a claim its numbers don't support. Several of the figures above also rest on this one blog, and the first MacRumors result was a single sample.

Now: the late-October window for 512GB has passed. These sources show neither a confirmed price nor a benchmark, so the first published run at that capacity is the next number that matters for anyone running models above 256GB.

Late January 2027: where Macworld said higher-end orders were already slipping at review time. Check the current ship date before planning around one. Results for the line sit on the LLM inference and Mac Studio pages.