The Mac mini M6 is out, and what the first benchmarks say about gaming and AI

Apple's Mac mini M6 went on sale on September 22, 2026 at $899 for 16GB and 256GB. That is $300 more than the M4 model it replaces, with no extra memory or storage for the money. The retail-unit reviews published since then agree on three things. Its CPU sits level with a 12-core M4 Pro. Its GPU lands just behind the 20-core M4 Pro in games. For local AI, the ceiling is the 153 to 170GB/s memory bus, not the chip.

These are tests on shipping units, not leaks. Tom's Guide published on September 21, 2026 and recorded 4,708 single-core and 21,405 multi-core in Geekbench 6. The multi-core figure is 44% above the base M4 and just under the M4 Pro's 22,661. Notebookcheck's chip analysis ran a 24GB/512GB unit through Geekbench 6.7 and Cinebench 2024. The rows below come from two different reviewers and should be read within each source, not across them.

Notebookcheck puts the M6's single-core result about 25% ahead of the M4 and about 10% ahead of the 10-core M5. Engadget, as cited by tbreak, measured it 16.5% ahead of the M4 Pro in Geekbench 6 single-core. In Cinebench 2026, Tom's Hardware ran the test ten times: 5,403 on the first run and 5,557 by the second. Gizmodo found the M6 18% ahead of a Ryzen AI 470 HX in Cinebench 2026 and 13% behind a Snapdragon X2 Elite Extreme. In Geekbench 7 it led a Core Ultra 9 386H by just over 20%. No Geekbench 7 score for the M6 itself appears in the reviews used here.

No review used here put the M6 against desktop x86 chips. The hub has Geekbench Browser averages dated September 16, 2026: 3,058 single and 26,572 multi for the Ryzen 9 9950X3D, and 2,832 and 25,837 for the Core Ultra 9 285K. Those are averages of user uploads, not one controlled run, so treat them as a rough reference. Even so, the M6 leads single-core by more than a methodology difference can explain, and both desktop chips are ahead on multi-core.

Gaming. Notebookcheck's numbers put the M6 GPU about 30% ahead of the 10-core M5 and about 80% ahead of the 10-core M4. It still trails the 20-core M4 Pro:

In a separate gaming piece dated September 29, Notebookcheck ran Cyberpunk 2077 at 1080p with ray tracing. It got 33 fps native, 44 fps with MetalFX upscaling and 80 fps with frame generation. Most of the jump to 80 is generated frames. Tom's Guide reported a stable 56 fps at 1080p on the "For this Mac" preset; the hub records the same review at 57 fps.

Real work. Gizmodo's Blender BMW CPU render took just over two minutes, about 40 seconds faster than the M4. The GPU render took 11 seconds against the M4's 36, though a Geekom AMD mini PC still finished about 15 seconds ahead. Tom's Guide found HandBrake noticeably faster than on the M4. tbreak says the M6 beat the M4 Pro in DaVinci Resolve and Pixelmator, but the facts here do not include its timings. No source here reports Xcode, Python or Lightroom Super Resolution times.

Local AI. StarryHope collected measurements from YouTubers and Reddit users. In llama.cpp, a 9B 4-bit model's prompt processing rose from 210 tokens/s on the M4 to 742 on the M6, and generation from 18 to 26.9. Alex Ziskind recorded time to first token falling from about 2.5s to 0.7s. Apple claims up to 4.8x faster LLM prompt processing over the M4. That vendor figure fits the prefill measurements but not the decode results, which improved much less.

Gizmodo's tenfold gap mostly comes from the M4 having less RAM, not from the chip. It shows what happens when a model does not fit in memory.

On efficiency, Notebookcheck measured 15.2 Cinebench 2024 points per watt single-core, against 13.5 for the M4 and 13.1 for the M5. Multi-core it fell to 32.5, behind the M4's 39. StarryHope saw a peak of about 38W during an LLM loop.

The M4 Mac mini launched in November 2024 at $599. The M6 replaces it at $899 with the same 16GB and 256GB, and every reviewer comments on it. Tom's Hardware's headline reads "Strong performance, but losing its grip on value". AppleInsider's says the hardware is "badly hampered by too-small SSD". Apple's own claims (up to 4x faster AI tasks and 2x faster graphics than the M4) are vendor figures for selected workloads, not standard benchmarks.

Every configuration has a 12-core CPU, a 12-core GPU and three 40Gb/s Thunderbolt 4 ports. Memory tops out at 32GB. For local AI, the 32GB model at $1,299 is the pick StarryHope and ComputingForGeeks both land on. Skip the 16GB unit for LLMs: StarryHope says a 27B 4-bit model would not even load on it, and Wired calls 16GB a practical ceiling for image generation. The reason is that macOS gives the GPU only part of the RAM; how much memory a Mac needs works through the arithmetic. Llama 3.3 70B fits on no M6 configuration.

Owners of an M4 Pro gain single-core speed and little else. The GPU and multi-core results are a wash, and the M4 Pro's 273GB/s bus generates tokens faster. M4 owners get the 44% multi-core jump and roughly 80% more GPU (Notebookcheck). If you need large models or want to game above 1080p, the Mac mini M5 Pro starts at $1,699 with 307GB/s and up to 64GB. Reviewers give it the gaming lead in the table above.

The sources here leave several gaps:

• Neural Engine core count, Wi-Fi and Bluetooth versions, Ethernet speed and the price of the 10GbE option.
• A Geekbench 7 score for the M6 itself, and Geekbench AI or Geekbench Metal results.
• tbreak's actual DaVinci Resolve and Pixelmator times.
• Xcode build and Python script times.
• Measured tokens/s for the M5 Pro Mac mini. StarryHope says none had been published by late September.
• A measured 32B result. Capital & Compute's 4.2 tok/s and ComputingForGeeks' 7 tok/s are both projections.

Notebookcheck expects the M6 to appear in the MacBook Air. Because that model has no fan, it expects somewhat lower performance than the Mac mini. Apple has not announced a date. The other open item is EveryMac's database, which still lists the new M6 models as "Pending" as of its November 2 update. When those Geekbench averages fill in, the hub's Geekbench 6 table will have a crowd-sourced number to set beside the review figures above.