Apple's M6 is a 2nm inference upgrade with a CPU refresh attached

Start with the comparisons Apple chose to publish. The M6-versus-M5 figures are specification deltas ^[2], while every number with an actual workload attached is measured against the M4 and the M1 ^[3]. Two of those sit in the same machine on the same baseline: LM Studio prompt processing up to 4.8x faster than M4, Excel calculations up to 1.5x faster than M4 ^[7]. The inference number is 3.2 times the spreadsheet number ^[1].

The spec sheet says the same thing more quietly. A 1.2x multithreaded CPU is a 20 percent uplift; a doubled Neural Engine is 100 percent, five times the improvement, and it lands on the part of the die that runs models ^[2][2]. Peak GPU compute for AI is up nearly 30 percent, unified memory bandwidth 10 percent ^[2]. Bandwidth is the number to watch there, because local model throughput tends to be fed by it, and 10 percent is not much of a meal.

The 2nm framing is the part most likely to get oversold at retail. TSMC puts N2 at 10 to 15 percent more speed at the same power, or 25 to 30 percent lower power at the same speed ^[9]. That comparison is against N3E, and the M5 is built on the later N3P, so MacRumors' read is that the real M5-to-M6 gain is smaller than the node headline implies ^[10]. Gate-all-around nanosheet transistors are a genuine change from the FinFET designs Apple has shipped for years ^[11]; they are not a change a buyer will feel in a spreadsheet.

The third core tier is the one architectural change that could show up in work with no model in it. Apple silicon has split cores into performance and efficiency groups since the M1; the M6 adds super cores for single-threaded work, and Apple claims the world's fastest single-threaded performance for them ^[12]. The workloads that cannot be spread across cores are the ones that benefit: compilation, file indexing, image editing, and the serial steps inside agentic workflows ^[13]. No one outside Apple has measured that claim yet.

So the upgrade case narrows to two groups. People whose bottleneck is one core, and people running models locally, where the doubled Neural Engine is exposed through system frameworks that can use both engines at once with no developer work ^[14]. For general responsiveness, MacRumors concludes the difference will be hard to notice and there is little reason to move from an M5 ^[15].

The lineup decision is the harder one. Bloomberg's Mark Gurman reported in July that Apple has no plans for M6 Pro, Max or Ultra parts, having pulled forward the AI-optimised M7 ^[4], and the higher-specification new Mac mini duly ships an M5 Pro ^[5]. The M7 is expected in the first half of 2027 with the neural upgrades that cut the M6 line short ^[16]. Which means the buyers with the strongest case for an M6 today are precisely the buyers the next chip is built for, and it is under a year out ^[16].