M6 MacBook Pro and iMac: 25% faster AI, OpenAI inside Pages

Apple refreshed the entry 14-inch MacBook Pro and the 24-inch iMac with M6 this morning, and the specs it led with are AI specs: up to 25 percent faster AI workloads than M5 on the laptop, a dual 16-core Neural Engine, and up to 24 hours of battery life, per Apple's announcement. Pre-orders open October 28, both machines arrive November 6, and the 14-inch still starts at $1,999.

The item in the same batch of releases that will matter longer has nothing to do with silicon. The new assistant inside Keynote, Pages and Numbers is "built on Apple models and OpenAI," and the same release adds Model Context Protocol support so outside AI agents can read a document, edit it, and run automated actions in all three apps, Apple says.

Put those two facts together and Apple's AI position is legible, and less flattering than the marketing: fast local silicon for the features Apple wants to own, someone else's model for the work people will pay for, and a document surface now open to agents Apple does not control.

What M6 actually changes

The chip is a wider version of the same design, not a new architecture. The M6 in the 14-inch MacBook Pro pairs a 12-core CPU — two super cores, four performance cores, six efficiency cores — with a 12-core GPU that now carries Neural Accelerators in every core, plus a dual 16-core Neural Engine that Apple says is up to 1.9 times faster than the single 16-core unit in M5. Memory bandwidth rises to 170 GB/s, 16 GB is standard and 32 GB is the ceiling. The N1 wireless chip that debuted in the iPhone 17 shows up here too, which is where Wi-Fi 7 and Bluetooth 6 come from, per 9to5Mac's comparison.

Apple's performance claims against M5: up to 30 percent faster graphics, up to 25 percent faster AI workloads, and up to 1.3 times faster time-to-first-token when processing an LLM prompt. The iMac gets the same silicon story with a bigger jump, because it skipped a generation — two more CPU cores than the M4 model, up to 12 CPU and 12 GPU cores, 170 GB/s, and up to twice the AI compute of the previous Neural Engine. Apple's own release also claims "up to 5x faster on-device AI performance than M4" in the opening paragraph, which is the sort of number that only makes sense once you know which workload it was measured on; treat the two figures as the same claim at different zoom levels. Prices start at $1,499 for the 10-core iMac and $1,699 for the 12-core.

9to5Mac's read is blunt and correct: those two chips are the entire update. Same chassis, same display, same ports. A redesigned high-end MacBook Pro is expected tomorrow, which tells you what today's machine is for — it is the value model, not the flagship.

The 1.3x is the number worth arguing about

Time-to-first-token is prefill, the phase where the GPU chews through your prompt before it writes a word. We made the case in September that prefill, not tokens per second, decides whether a local agent is usable at all, because an agent loop resends its system prompt, tool descriptions and session memory on every turn — M5 Ultra reviews: local AI's ceiling is memory, not compute. A 30 percent faster prefill makes that loop feel less broken. It does not make a bigger model fit.

Two numbers bound what this laptop can do locally. The first is 32 GB, the maximum unified memory in the 14-inch. A 30-billion-parameter model at four bits per weight lands near 17 GB before you add a context window, so a machine like this runs a mid-size model comfortably and a frontier-scale one not at all. The second is 170 GB/s of bandwidth, which is the hard ceiling on how fast the GPU can stream weights during generation — the M5 Ultra Mac Studio we covered in September moves 1.2 TB per second, roughly seven times more. That gap is why the same company sells both a $1,999 laptop and a desktop with 1.2 TB per second of memory bandwidth to people running models locally.

So the honest framing of a 25 percent AI gain: it is a real improvement in how quickly the machine responds, and no improvement at all in what it can hold.

Apple concedes the model layer

The Creator Studio assistant is the interesting admission. Apple has spent two years selling Apple Intelligence as the privacy-safe alternative to the cloud, and its flagship productivity assistant is partly OpenAI's. Apple has not published which requests run on the Neural Engine, which go to its private cloud, and which reach OpenAI — that split is undisclosed, and it is the first thing a skeptical buyer should ask about.

The MCP support is the strategically bigger move, and it barely got a line in the release. Keynote, Pages and Numbers become tool surfaces: an outside agent — Claude, ChatGPT, whatever a team has wired up — can read the deck, restructure the document, act on the spreadsheet. If you want the background on why that matters, AI 101 — What is MCP? covers the protocol and Apple's new Mac Studio and Mac mini are built for local AI covers the hardware bet underneath it.

Read the strategy plainly and the roles are clear. Apple keeps the surface where the work lives — the document, the deck, the spreadsheet — and the subscription that reaches it. OpenAI gets distribution inside the default word processor on every Mac, iPhone and iPad that installs the update. Local model runners get faster hardware and an unchanged memory ceiling. The party that loses is the one that assumed Apple would eventually ship its own frontier model and route everything through it; that assumption is now three product cycles old and still wrong.

What the skeptics should press on

The assistant is in beta, and it requires iOS 27, iPadOS 27 or macOS 27 — so the installed base that can use it on day one is the part that has already updated, per The Verge's write-up. The performance figures are Apple's, measured against M5 configurations Apple chose, and "up to" is doing its usual work in every sentence. A spec bump that lands the day before a redesign is also a strange thing to buy on day one.

None of that makes M6 a bad chip. It makes it an incremental one, shipped into a year when the interesting AI work moved to datacenters financed with tens of billions in debt. Apple's answer to that is not a bigger model. It is a faster Neural Engine, a 32 GB ceiling, and an OpenAI key it does not advertise.

What to watch

Whether Apple publishes the on-device versus cloud split for the Creator Studio assistant, because a privacy claim you cannot audit is a marketing claim. Whether third-party agents actually connect to Pages through MCP, or whether the support ships and sits unused like most new protocol surfaces in their first year. Whether tomorrow's redesigned MacBook Pro raises the memory ceiling, which would matter more to anyone running models locally than another 25 percent. And whether the iMac's 5x-on-device-AI line survives contact with a benchmark that is not Apple's.

If the assistant in your word processor is partly someone else's model, do you want to know which requests leave the machine? Write to us via our contact page.