Apple's A20 Pro is the first 2nm phone chip, tuned for on-device AI
A doubled 32-core Neural Engine and 2x FP8 throughput let iPhone 18 Pro run third-party LLMs locally. Plus: ChatGPT Voice moves to GPT-6 Astra with GPT-Live.
Copy markdown
A desktop-class NPU for local models
The 32-core Neural Engine — now two engines — delivers 2x the compute of the A19 Pro, and the GPU's neural accelerators double FP8 throughput. Apple pitched the chip explicitly for “running third-party LLMs on-device,” which means larger local models and faster tokens with no API round-trip.
The bottleneck it fixes is bandwidth
Local LLM decoding is memory-bound, and the A20 Pro ships “the widest memory interface ever in an iPhone” — a 50% bandwidth jump — with the memory die moved off the thermal path onto a new vapor chamber. Sustained performance is up 40% over the A19 Pro, so long agent runs throttle less.
First 2nm silicon, only in iPhone 18 Pro
It's the first 2nm smartphone chip: a 6-core CPU with two “supercores” ~20% faster and a 7-core GPU up to 40% faster. It ships only in the iPhone 18 Pro and Pro Max, so the on-device gains land when those units reach hands — not a download you can grab today.
Elsewhere: ChatGPT Voice moves to GPT-6 Astra
OpenAI put GPT-5.6 and GPT-6 Astra behind ChatGPT Voice and introduced GPT-Live at $0.05/min (or 1 credit/min), retiring the old Instant/Medium/High voice levels and cutting extra-voice cost from 5 credits to 1. If you lean on Voice, re-check your model routing and per-minute spend.