Generating images, video and voice on a Mac against the best AI APIs (October 2026)

Which of these came from a laptop?

All eight pictures were made from the same sentence. Two came from an M1 Max MacBook Pro running open models, and six from hosted APIs that cost between one and fifty cents a picture. Have a guess before you press the button.

“Two friends laughing while sharing a pizza at a kitchen table, candid photo, natural window light, hands visible, slight motion blur”

Alocal

Z-Image Turbomade on the M1 Max in 1 min 38 s, $0.0003 of power

Bhosted

Muse Image (Meta)$0.01 an image

Chosted

GPT Image 2.5 Sunburst (max)$0.50 an image

Dlocal

FLUX 3 Image Kleinmade on the M1 Max in 1 min 45 s, $0.0004 of power

Ehosted

Nano Banana 2.1$0.034 an image

Fhosted

Seedream 5 Pro$0.045 an image

Ghosted

FLUX 3 Image$0.024 an image

Hhosted

FLUX 1.1 pro$0.04 an image

Show which is which

For everyday pictures like this one, the difference between a laptop and the best hosted models has become hard to see. The rest of this post is about where it still shows, and what each side costs.

Six prompts, fourteen image models

We wrote six prompts to test different things: a close-up portrait, an interior, a chalkboard with hand-lettered text, two people at a kitchen table, a real place (Helsinki harbour in winter) and a flat illustration. Three models ran on the M1 Max and eleven through Replicate and OpenRouter. Click any picture to see it full size, use the buttons to show only one side, or switch to By model to see all six pictures from one model at a glance.

By promptBy model

portraitofficechalkboardfriendsharbourrobot

AllOn the MacHosted

“Close-up portrait of an elderly fisherman on a wooden pier, weathered skin, knitted wool cap, overcast soft light, shot on an 85mm lens, shallow depth of field, natural colours”

local

Z-Image TurboM1 Max · 1 min 48 s · $0.0004 of power · Elo 936

local

FLUX 3 Image KleinM1 Max · 1 min 46 s · $0.0004 of power · Elo 1043

local

FLUX.2 klein 4BM1 Max · 26 s · $9e-05 of power · Elo 837

hosted

GPT Image 2.5 Sunburst (max)2 min 48 s · $0.50 · Elo 1222

hosted

GPT Image 2.5 Flare (max)63 s · $0.50 · Elo 1207

hosted

Nano Banana 2.116 s · $0.034 · Elo 1176

hosted

MAI-Image-2.627 s · ~$0.04 · Elo 1166

hosted

Grok Imagine Image 254 s · $0.04 · Elo 1166

hosted

Muse Image (Meta)26 s · $0.01 · Elo 1127

hosted

MAI-Image-2.6-Flash12 s · ~$0.02 · Elo 1119

hosted

FLUX 3 Image20 s · $0.024 · Elo 1117

hosted

Seedream 5 Pro41 s · $0.045 · Elo 1100

hosted

Qwen-Image 3 Pro57 s · $0.04 · Elo 1095

hosted

FLUX 1.1 pro3 s · $0.04 · Elo 870

Every model, local or hosted, wrote the chalkboard sentence correctly, which not long ago was the thing image models were worst at. The hosted models pull ahead on real places: Nano Banana drew a Helsinki you can recognise, while the local models and several of the hosted ones drew a generic harbour. FLUX 1.1 pro, which was the default in our own tooling until this test, looks dated next to the rest, with the over-smoothed skin people now recognise as an AI picture. On the Mac, Z-Image Turbo looks the most like a photograph, FLUX 3 Image Klein has the best texture and lettering, and FLUX.2 klein 4B is four times faster than either and loses detail for it.

What a picture costs, and how long it takes

A picture made on the Mac costs about $0.0003 of electricity; we measured 39 W while generating against 2 W idle. If you also count the Mac's own time, say a $2,000 laptop over four years of generating eight hours a night, it comes to about half a cent. The cheapest hosted model we tried, Meta's Muse Image, costs one cent, and the most expensive, GPT Image 2.5 at maximum quality, fifty.

Price and rating barely move together. Muse at one cent is rated above models that cost two to four times as much, and GPT Image 2.5's lead over Nano Banana costs fifteen times as much per picture. The green dots are the Mac: FLUX 3 Image Klein is the highest-rated open model on the leaderboard, while Z-Image Turbo sits well below the hosted models even though, to our eyes, it made the most photographic pictures of the three.

On time the Mac is in the same range as the hosted models, from 26 seconds to under two minutes a picture. What a hosted API gives you is the ability to make a hundred at once, where the Mac makes them one after another, so a big batch on the Mac is something you start before bed.

Video

Video is where we expected the Mac to fall furthest behind. Three prompts: a barista pouring latte art, a ferry leaving a snowy harbour, and a dog catching a ball. Three models ran on the M1 Max and eight hosted, the hosted ones at 5 seconds and 720p with sound. Every clip shows how long it took and what it cost, and By model shows all three clips from one model.

The Mac is behind on resolution and length rather than on time. LTX-2.5 makes four seconds at 768×512 with sound in about two and a half minutes, which is about what most of the hosted models took for five seconds at 720p. Wan 2.2 got the physics of the latte pour right where LTX did not (the milk comes from the pitcher), but at over an hour a clip it is something you start at night.

Hosted, a clip cost from 9 cents (HeyGen Video 1, measured) to $1.68 (Kling 3 Pro). We have only judged the hosted clips from still frames so far, so watch them and make up your own mind about motion and sound.

The same people, twelve years apart

For our own testing we need a believable photo library of people who do not exist: the same family across years of breakfasts, playgrounds and climbing trips. That needs a model that keeps a face recognisable from one picture to the next. We gave each model portraits of the people in the scene as reference images, one per person, and named what each of them wore. The cast comes from open-synthetic-identities.

alex, age 26

alex, age 38

sam, age 32

sam, canonical

mira, age 2

mira, age 7

jonas, age 26

breakfast 2026park 2021climbing 2015

References: alex/age-38, sam/canonical, mira/age-7. “Candid smartphone snapshot from 2026: the adult in image 1 (wearing a navy hoodie) and the adult in image 2 (wearing a plain white t-shirt) eat breakfast at a small kitchen table with the child in image 3 (wearing yellow pyjamas), a Saturday morning in a Berlin apartment, cereal bowls and orange juice, the child laughing, slightly tilted framing, mixed window light, a little motion blur, an ordinary phone photo, not a studio portrait”

local

FLUX 3 Image KleinM1 Max · 11 min · $0.002 of power

hosted

GPT Image 2.51 min 50 s · $0.50 a picture

hosted

Nano Banana 2.135 s · $0.03 a picture

hosted

FLUX 3 Image1 min 35 s · $0.02 a picture

hosted

Seedream 5 Pro1 min 52 s · $0.04 a picture

hosted

Muse Image17 s · $0.01 a picture

hosted

refused: a child in the picture

MAI-Image 2.6no picture

hosted

refused: a child in the picture

MAI-Image 2.6 Flashno picture

Every hosted model that answered kept the faces, and so did FLUX 3 Image Klein on the Mac, through mflux's edit command, which takes several reference images; it needed 8 to 11 minutes a picture against two minutes at most hosted. Two things to know if you try this. Name the clothes, or the model copies the outfit from the reference portrait and everybody wears the same thing for twelve years. And MAI-Image 2.6 refuses every scene with a child in it, while Muse refused the climbing scene, which has none, without saying why.

Voices

Last, speech. We wrote three short voice memos, one in English and two in German, described each speaker's voice in a sentence, and had every model read them. Models that can design a voice from a description got the description, and the others their closest preset. To check each take we transcribed it with whisper and counted the words it got wrong; a few percent is whisper's own error.

For reading a script, the Mac is as good as the hosted services: on the Mac, Ming-3.0, Qwen3-TTS and Chatterbox did not get a word wrong in any memo, and Ming-3.0 needs only 1.5 GB of memory. Two takes went wrong in ways worth knowing about. When we put the voice description in the same paragraph as the text, Gemini read the description out loud before the memo (the last one in the English list); putting it under its own heading fixed it. And Fish Audio Drama 3 stopped its English memo after two sentences and filled the rest with noise. How natural each voice sounds is a matter of taste, which is why they are all here to play.

How we ran it

The local models ran on an M1 Max MacBook Pro with 64 GB through their MLX ports (mflux 0.23 for images, ltx-2-mlx and mlx-video for video, mlx-audio 0.6.1 for speech), with seed 42, mostly overnight on 24 to 25 October. The hosted models ran through Replicate and OpenRouter on 26 October at list price, and the ratings are from the Artificial Analysis leaderboards read the same day (the Mac models' on 28 October). Licences differ: Z-Image Turbo, FLUX.2 klein 4B and Wan 2.2 are Apache-2.0; FLUX 3 Image Klein is free under US$5M revenue and LTX under US$10M a year. The scripts, prompts, every output and the traps we hit getting the local models to run are in the october-2026 folder of the benchmark repo. For how fast the same Mac runs language models, see the companion post.