TodayToday's edition
Four reviews of Apple's $1,999 M6 MacBook Pro, from three publishers, find gains over M5. Xataka's local-AI test hit 25.7 tokens/s, but the 16GB unit could not load gpt-oss-20b.
Hardware & Analysis DeskEdited by Ege Erten
4 min readAI-assisted
Follows our story: Apple MacBook Pro M6 Tests: Geekbench GPU 64,115 vs 48,840 on M5
Apple's 14-inch MacBook Pro with M6, from Apple's product specifications page.
Image: Apple
Source · Press image from the Apple website, used for news reporting.
In reviews of Apple's M6 MacBook Pro, Xataka's local language-model test, the first Voltgrid has found, measured 25.7 tokens per second from Qwen 3.5 9B on the 16GB base unit, which could not load gpt-oss-20b.
That result fills the gap left in Voltgrid's Nov. 5 report on the first Tom's Hardware and Engadget numbers, when Voltgrid had found no outlet that had run a model on the laptop. Four outlets have now published measured results, all dated Nov. 5, but they come from three publishers. Xataka and Applesfera are separate newsrooms owned by the same Spanish publisher, Webedia. Tom's Hardware belongs to Future, whose Tom's Guide prints the same lab results and is not counted again. Xataka, Tom's Hardware and Engadget reviewed 16GB, 1TB units, the $1,999 U.S. base configuration (Xataka priced its Apple-loaned unit at 2,329 euros); Applesfera did not state its configuration.
Apple says the M6 is 20 to 30 percent faster than the M5 depending on the application, according to Notebookcheck's launch report, and claims up to 30% faster graphics and 25% faster AI. Xataka points out that Apple's 24-hour battery figure comes from a 32GB machine. Apple's release says up to 170GB/s; per Xataka, the 16GB model runs at 153GB/s, the same as the M5, while the 24GB and 32GB models run at 170GB/s, as set out in Voltgrid's M6 spec file.
On the CPU, Tom's Hardware measured Geekbench 6 at 4,752 single-core against 4,288 on its M5 unit, about 10% more, and a multi-core gain of just over 18%. Engadget's run gave 4,458 and 19,077 against 4,310 and 18,003 on its M5, and about 30% more than M5 in Cinebench R24 multi-threaded work. Xataka used the newer Geekbench 7.1, scoring 3,966 and 20,049, which cannot be set against Geekbench 6 results. In Cinebench 2026 it measured 802 single-thread and 5,384 multi-thread points.
Graphics moved more than the CPU. Engadget's Geekbench 6 GPU score rose to 64,115 from 48,840, and Xataka recorded 88,303 in Geekbench 7.1 Metal. Tom's Hardware measured Cyberpunk 2077 at 53.7 fps at 1,800 by 1,125, against 47.65 fps on M5, and timed a HandBrake 4K-to-1080p transcode at 2:50 against 3:31. Applesfera measured 54 fps decoding 8K Blackmagic RAW on the CPU and 130 fps through Metal, and 335.68 points in Blender 5.2.0's CPU test.
Storage barely moved. Xataka measured 6,660MB/s read and 6,551MB/s write, and Applesfera 6,701.7MB/s and 6,447MB/s in Blackmagic Disk Speed Test. Tom's Hardware's 25GB file copy ran at 1,914.23MBps, against 1,917.35MBps on M5.
A language model writes one token at a time and reads its weights from memory for each one, so bandwidth sets how fast it replies and capacity sets which models fit. Xataka reckons about 12GB to 13GB of the 16GB is left for the model and its context. Summarizing a 10,000-token article in LM Studio, it reported 25.7 tokens per second from Qwen 3.5 9B, 15.5 from Gemma 4 12B QAT and 36.2 from NVIDIA's Nemotron 3 Nano 4B. Xataka warns that the models ran in different formats, MLX and GGUF, so the table does not compare size alone.
Speed held up over time. Across 20 back-to-back Qwen runs, Xataka saw 27.9 tokens per second at the start and 28.6 at the end, with two dips to about 22 in between, and it did not hear the fan with small models. Larger models failed outright: the 12.1GB gpt-oss-20b and Qwen 3.8 27B would not load, and an agent task with the 9B model used 14.01GB of 16GB plus 1.81GB of swap.
Apple's AI figure is a vendor claim and Xataka's is a measured result on a 16GB unit, so they are different measures. Apple describes its figure as covering AI workloads such as prompt processing, the step that reads the input; Xataka does not say whether its tokens-per-second figure covers generation alone or includes prompt processing. Voltgrid has found no outlet that has run a model on a 24GB or 32GB unit at 170GB/s.
Tom's Hardware and Engadget tested the same configuration but measured 4,752 and 4,458 points in Geekbench 6 single-core, and neither explains the gap. The figures should not be averaged.
Battery results reflect different workloads, not a conflict. Tom's Hardware measured 19 hours 5 minutes of web browsing, OpenGL tests and video at 150 nits, nearly an hour more than its M5. Engadget recorded 9 hours 35 minutes in a script that simulates programming. Xataka got just under 13 hours of mixed real work with no fixed method. Apple's 24-hour figure is 1080p streaming at low brightness on a 32GB unit, Xataka reports.
On heat, Tom's Hardware's ten Cinebench 2026 runs opened at 5,026.55 and held around 5,440, with the chip peaking at 80°C, though its tool does not yet fully support M6. Xataka found the fan very loud in multi-threaded Cinebench. It planned a 30-minute sustained test that logged watts but did not finish it, so Voltgrid found no published power draw for the chip.
Memory is the spending decision. Apple charges $200 more for 24GB and $400 more for 32GB, which also lifts bandwidth to 170GB/s. Engadget prices the M5 Pro model at $2,499, with up to 64GB at 307GB/s. Xataka recommends 24GB for anyone who runs models locally with some regularity, and Applesfera calls 24GB the balanced choice.
Every outlet measured gains over the M5 in CPU and graphics, with storage flat, from what Engadget calls a slight upgrade to clearer steps in graphics. For local AI, the 16GB base model suits small models such as a 9B at about 25 tokens per second, a pace Xataka calls comfortable to read, but it rules out 20B-class models such as gpt-oss-20b. Xataka and Applesfera point video, 3D and large-model work to the M5 Pro, and Xataka advises M5 owners not to upgrade. Voltgrid found no published local-model run on a 170GB/s unit, no power figure under load and no full Notebookcheck laptop review.
ChipsReview roundup
Local AIComparison
