TestFlight beta iPhone and iPad · iOS 18 or later

The AI benchmark that makes your iPhone sweat.

RapidBench is an AI benchmark that makes your iPhone sweat: four real AI workloads, run on the phone with nothing sent to a server. Same models, same prompts, measured the same way, so the only thing left to compare is the chip, and one score you can put next to anyone’s.

iPhone 18 Pro vs 17 Pro.

The full benchmark on both phones, side by side. The A20 Pro gets through on-device AI about 50% faster.

iPhone 18 Pro and iPhone 17 Pro running the RapidBench Neural Engine test side by side

Measured on iPhone 18 Pro, 22 September 2026. Slowed down 10× so you can watch it happen. Chip details from Apple’s A20 Pro and M6 announcements.

One Neural Engine, or two?

The Neural Engine is the part of the chip built only for AI. It can’t draw graphics or run apps, but for the math AI models are made of, it’s fast and sips power. It’s what the iPhone leans on for photo processing, Face ID and on-device AI models.

Until this year every iPhone had one. The A20 Pro, in the iPhone 18 Pro, 18 Pro Max and iPhone Duo, has two: 32 cores where the A19 Pro had 16. Apple’s M6 Macs got the same dual design. Two engines can each take a job at the same time, so a test that only ever sends one would miss half the hardware.

So RapidBench sends the work down one lane, then two at once: a 1024 × 1024 photo cut into 25 tiles, each sharpened to twice the size. On a chip with one engine, the second lane just waits its turn. On the A20 Pro it runs alongside the first, 145 tiles a second instead of 85. The GPU and CPU then run the same work, for comparison.

tiles per second · pick one to run it

Pegged to a real phone.

Scores are pegged against an iPhone 18 Pro, which scores around 5,000. A number only means something next to another number, so we picked a real phone you can picture and made it the middle of the scale.

Points grow in a straight line with speed. Half as fast is 2,500, twice as fast is 10,000, and there’s no ceiling. Only speed earns points: a wrong answer or a blurrier picture can take points away, never add them. Memory and heat sit next to the score instead of being mixed into it, so you can see them without them muddying the number.

Your phone 1.00× the speed
5,000

points = 5,000 × your speed ÷ the iPhone 18 Pro’s

Inference

  • Decode speed, first 5 prompts40
  • Decode speed, last 5 prompts35
  • Median time to first token25

Lowered by wrong GSM8K answers. Badges: peak memory, and how much speed holds.

Training

  • Steps per second60
  • Total time, load and evals included40

Lowered by a worse validation loss. Badge: peak memory.

Image

  • Images per minute60
  • Denoising steps per second40

Badge: peak memory.

Neural Engine

  • Tiles per second, best lane count60
  • Single-request latency40

Lowered if the upscale gains less sharpness over plain resizing.

Each test is the weighted geometric mean of its speed metrics; the overall score is the geometric mean of the four. Quality can lower a score to half, never raise it. Reports keep raw numbers, so old runs are rescored when the reference moves.

Thermal impact

Phones get hot, and a hot chip slows itself down to protect the battery and your hand. This is the same model answering the same question on an iPhone 18 Pro, replayed at the speeds it measured: about 86 tokens a second while cool, and about 40 once eleven questions have warmed it up. Flip between them to feel the difference.

That’s why the score times the first five questions and the last five separately. A phone that keeps its speed when it’s warm scores better than one that only sprints.

GSM8K · problem 1 of 20 Cool chip

Janet’s ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers’ market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers’ market?

First token
–
Tokens
0
Elapsed
000 s
Decode
– tok/s

Pace from the reference run on 22 September 2026. The wording of the reply is ours; the speed is the phone’s.

We sweated the small stuff too.

Downloads that don’t need you.

The models download in the background, even with the phone locked. A Live Activity follows them in the Dynamic Island and on the Lock Screen, and tells you when they’re ready.

64 GB

of models, fetched once, resumed if the Wi-Fi drops, and kept out of your iCloud backup.

0 B

of your prompts, pictures and reports leave the phone. Scores go to Game Center only if you’re signed in.

It waits for the phone to cool.

A phone that’s already warm waits up to five minutes before the suite starts, so no run begins throttled. In a hurry? There’s a Skip button.

Every run, kept.

Reports save the raw numbers and the images, not just the score.

  • ri1.overall
  • ri1.inference
  • ri1.training
  • ri1.image
  • ri1.ane

Ranked by phone, too.

Five Game Center boards. Every score carries its chip, so phones get ranked, not just players.

Same models · same data

No asterisks.

Every phone runs the same models on the same data: the same 20 math problems, the same 500 training rows, the same three prompts and seeds, the same photo. Nothing bends to fit a phone, so what you see is the chip’s real speed.

Put your iPhone on the bench.

RapidBench is in beta on TestFlight. Drop your email and get the TestFlight link!

  • iOS 18 or later
  • Neural Engine test needs iOS 27
  • About 6.4 GB of models
  • Wi-Fi recommended