MacBook SVG Benchmark: 26 AI Models, Ranked

The first model to take the top seat off the Fables, and it does it on the criterion this page actually grades: assembly. All four of its cells - even default at $0.33 - draw a laptop whose lid and base share a hinge and a vanishing point, with nothing floating, torn, or overhanging; no other ladder in the catalog manages that. At max it renders what looks like a product shot: an Apple-style menu bar with status icons and the 9:41 clock, a dock of individually recognisable apps, a notch, side ports, speaker grilles, a full keyboard with function row - 45.8K tokens, 13.8 minutes, $2.29, roughly half the tokens, time and money of Fable 5.1's dethroned best. The pelican at max is a children's-book illustration: saddle, bell, brake lever, fenders, a chain that actually wraps the sprocket. The effort dial is cleanly ordered on both tasks - the first frontier model here where paying more reliably buys more. OpenAI launched it behind a limited-access program on 2026-09-04; our access opened the next morning and it went straight through smoke, tokenizer verification (o200k, ratio 1.0000) and this ladder in the same hour.

• default - Perfect assembly out of the gate: joined hinge, one vanishing point, keyboard, trackpad, ports - no menu bar or dock yet, but nothing wrong.
• xhigh - A dock with a badge on the Mail icon, a dated menu bar, ports and grilles - $0.98 for a drawing that would have led this page a week ago.
• max - The new best drawing in the benchmark: an Apple-style menu bar with status icons and the 9:41 clock, a dock of recognisable apps, notch, side ports, full keyboard - and not one visible assembly error. Half the tokens, time and price of the Fable 5.1 cell it dethrones.

In Playcode since 2026-09-05 · runs 2026-09-05

The GPT-6 price cut lands on the benchmark the same day it lands on the API, and it rearranges the top of this page on value rather than on quality. Its max cell is a genuine product shot - an Apple menu bar with named menus, wifi and battery glyphs and the 9:41 clock, a dock of individually recognisable apps, a notch, perforated grilles, a trackpad with its lip, and a full keyboard where the letter, modifier and function keys all carry legible legends - and it costs $0.386 against the $2.29 GPT-6 Astra spent on the drawing that still leads. Astra keeps rank 1 by a hair, on side ports Sol does not draw. What makes this more than one good cell is the ladder: all four assemble correctly, including the cheapest at 3.6 cents, which still has a notch, a real key grid, grilles, a trackpad and a lid and base that share one vanishing point. Nothing floats, tears or overhangs anywhere in the set, and the dial is cleanly ordered on both tasks - 3.6K tokens at default rising to 38.5K at max. Its pelican is a tidy, correct illustration at every effort. Half the price of GPT-5.6 Sol, which it replaces on the Playcode Pro tier, and eleven places above it here.

• default - Three and a half cents, 48 seconds, and nothing wrong with it: notch and camera, a real key grid, perforated grilles both sides, a trackpad, and a lid and base that share one vanishing point. No menu bar or dock - the screen is bare wallpaper - which is the only thing separating this from cells costing fifty times more.
• high - The furniture arrives: an Apple menu bar with Finder, File, Edit and View, a wifi glyph and the 9:41 clock, a dock of coloured app tiles over a mountain wallpaper. Body, keyboard, grilles and trackpad all still correct.
• xhigh - A fuller menu bar (through Window and Help), rounded dock icons with distinguishable glyphs, and keys that begin to carry legends. The only blemish is the Help menu label running into the wallpaper behind it.
• max - The product shot, for 39 cents. A full Finder menu bar with wifi and battery glyphs and the Mon 9:41 clock; a dock of individually recognisable apps; a notch; perforated grilles; a trackpad with its lip; and a keyboard whose letter, modifier and function keys are all legibly labelled. Only side ports are missing against the drawing that still leads this page at six times the price.

In Playcode since 2026-09-22 · runs 2026-09-23

Released 2026-09-21 and benchmarked the next day, and it rearranges the top of this page without taking it. Its xhigh cell is a product shot: an Apple menu bar with named menus, status icons and the 9:41 clock, a dock of individually recognisable apps, a notch, side ports, speaker grilles, the MacBook Pro wordmark on the chin, a keyboard that recedes with the deck, and one vanishing point shared by lid and base. That is detail parity with the rank-1 Astra drawing - it even has the chin wordmark Astra omits - and it costs $1.72 against Astra's $2.29 and a third of the $4.95 Fable 5.1 cell it matches. Astra keeps the top seat on two things: legible legends on every key, and a ladder where all four cells land. Cheap is where Opus 5.5 is most convincing: high is $0.40 for a correctly assembled laptop with grilles, a real key grid and a trackpad - the cell Opus 5 gets faulted for, drawn right - at 21% less cost and 32% less time than Opus 5 spent on its near-identical version, which is Anthropic's claimed 30% speed-up showing up in an independent measurement. The ladder then breaks at the top, and expensively. At max it spent all 128,000 output tokens - its own ceiling, not our sweep budget - on thinking, emitted nothing until the stream closed, and billed $2.56 for an empty file. Fable 5.1 failed this way against a 64K budget and finished when given its real ceiling; Opus 5.5 had its real ceiling from the start, so there is no larger number to re-run it with. The pelican dial is ordered and cheap: $0.05 buys a complete bird and bike with no saddle and a stub of a chain, $0.13 buys a finished illustration - chain wrapped over both sprockets, cranks, legs on the pedals, motion lines, sun. A postscript on that empty max cell: it was never a runaway. Re-run on the Batch API, the one surface that allows this model more than 128,000 output tokens, the identical request finished naturally at 255,957 - and drew the most complete laptop in this catalog, menu bar and clock and dock and side ports and labelled keys and glow on the desk. The model is not failing the task at max effort; the streaming API simply cannot hold the answer. That drawing costs $2.56 at the Batch API's half price, arrives asynchronously hours later, and is therefore not something you can ask for in a chat - which is why it sits beside the empty cell rather than replacing it.

• high - Forty cents for a laptop with nothing wrong with it: notch, menu bar, a dock of app tiles, perforated grilles both sides, a real key grid and a trackpad, lid and deck at one eye height. Opus 5 spent 20,145 tokens and 256 seconds on the same drawing and got a skewed base; this is 19,767 and 173.
• xhigh - The product shot, and the best drawing on this page under $2. A real Finder menu bar with named menus, status icons and the 9:41 clock; a dock of individually recognisable apps; notch, side ports, grilles, the MacBook Pro wordmark, a keyboard that recedes with the deck. 86,096 output tokens against 28.5 KB of SVG - about 57K of it thinking.
• max - Spent all 128,000 output tokens - the whole ceiling the streaming API allows this model - on thinking, and returned an empty file. Nothing arrived until the stream closed 20 minutes in. $2.56 for no drawing. Re-running it on the same API cannot help: 128,001 is rejected by name. The row below is the same request given 300,000 tokens on the Batch API, and it needed 255,957 of them.
• max - DIFFERENT METHOD, and the only cell on this page that is: the Batch API with the output-300k-2026-03-24 beta, which is the one surface that allows this model more than 128,000 output tokens. It finished naturally at 255,957 - twice what the streaming API permits - and it is arguably the best drawing in the catalog: a full Apple menu bar with status icons and the Tue 9:41 clock, a dock of individually recognisable apps, a notch, side ports, finely perforated grilles, a legibly-labelled keyboard with function row, and screen glow bleeding colour onto the desk. Billed at the Batch API's 50% discount ($2.56; the same tokens on the standard card would be $5.12), asynchronously, over about two and a half hours - so it is not something you can ask for in a chat.

In Playcode since 2026-09-22 · runs 2026-09-22, 2026-09-23

Anthropic's newest model, and the one that broke this benchmark's budget rule. On launch day its xhigh and max cells returned nothing: thinking is always on for 5.1, and both spent our whole 64K sweep budget reasoning without emitting a byte. Re-run at the model's own 128K ceiling, both finished - and max is the best drawing on this page. Not an isometric diagram like Fable 5's: a render in real perspective, with a notch, a menu bar, a dock, the MacBook Pro wordmark, speaker grilles, three side ports, a keyboard that recedes correctly and screen glow spilling onto the desk. The only visible error is a trackpad sitting slightly right of the keyboard centre. It costs $4.95 and 19 minutes; xhigh is $4.34 for a thicker, slabbier body seen from the other side; high is $0.90 for most of the same laptop with the base a few degrees off the lid. The first sweep's 16K cap is a wall for it - thinking alone exhausts it - which is why the cheap tier is not where you use this model. The pelican ladder finished at every effort: at high the head floats free of the body, at xhigh it is a proper bird leaning over the bars, and at max it is the cleanest pelican in the benchmark for $1.92.

• high - Uncapped (64K): silver body, notch, menu bar, dock, grilles, a perspective keyboard - and a body the checklist takes apart. The base is a wedge rather than a slab, it ends in a spike at the front-right corner, and its footprint does not match the lid. Read by eye this passed for "a few degrees of skew"; three separate checklist runs all named the wedge and the spike.
• xhigh - At the model ceiling (128K): finished in 17.2 minutes and 86,716 tokens. The same silver three-quarter body seen from the other side, with a menu bar, a dock and side ports - but a thick, slab-like base and a wide empty palm rest. The 5.1 cell that changed the benchmark: at our old 64K budget this run drew nothing.
• max - At the model ceiling (128K): 99,072 tokens, 19 minutes, $4.95 - and the best drawing in the benchmark. Perspective, not isometry: notch, menu bar, dock, wordmark, grilles, three side ports, a keyboard in perspective and screen glow spilling onto the desk. The trackpad sits slightly right of the keyboard centre; nothing else is visibly wrong.
• high @16K - At the first sweep's 16K cap: thinking took almost all of it and 1.5 KB of SVG was cut off mid-tag. Fable 5 finished xhigh inside the same cap.
• xhigh @64K - Cut off by OUR 64K sweep budget: 12.4 minutes of thinking, not one byte of SVG, $3.20 billed. Re-run at the model ceiling it finished - the wall was ours.
• xhigh @16K - At the 16K cap with zero visible output - the entire budget went to thinking. On 5.1 the old cap is not a budget, it is a wall.
• max @64K - Cut off by OUR 64K sweep budget: 64,000 tokens of thinking in 12.4 minutes, no SVG, $3.20 for nothing. Given its own 128K ceiling the same request finished in 99,072 tokens.

In Playcode since 2026-09-01 · runs 2026-09-01, 2026-09-02

The most careful draughtsman here. Its xhigh MacBook won the launch sweep, and at max it draws the only laptop on this page with no visible error at all - notch, dock, grilles, side ports, every edge meeting where it should. What keeps it second is the projection: it draws a flat isometric diagram, where the prompt asks for a perspective render with screen glow, and 5.1 delivers exactly that. $2.13 and nearly ten minutes for one laptop; half the price of its successor's best cell, and half the ambition.

• xhigh - Winner of the launch sweep - the most detailed MacBook anyone drew inside the 16K cap.
• max - Uncapped (64K, streamed): the best drawing in the benchmark - isometric, dock, side ports - for $2.13 and 9.8 minutes.

In Playcode since 2026-07-10 · runs 2026-07-13, 2026-07-21

Superseded by GPT-6 Sol, which costs half as much and ranks fifteen places higher - see rank 2. The most ambitious - more gradients, more perspective, more parts than anyone. At high that ambition bends the geometry; at xhigh it pulls together into a genuinely good laptop, admittedly more dark Lenovo than MacBook. Its party trick is the failure: it finishes the pelican at max, but on the MacBook it never returns at all.

• max - Timed out. It did not return in 36 minutes, even with a 64K budget - exactly as in the first run.

In Playcode since 2026-07-09 · runs 2026-07-13, 2026-07-21

The most detailed, ambitious Anthropic laptop here - notch, dock, traffic lights, a code window, convincing materials - and the first Anthropic model whose effort dial actually changes the drawing instead of redrawing the same thing. But it fails the one thing the MacBook test is for: the geometry. At high the base is skewed; at xhigh the isometry is broken outright - the screen and the keyboard deck sit in different perspectives, so it reads slick at a glance and falls apart on a second look. And max was the worst of the three until we found the cause: thinking is on by default, so it spent almost our entire 64K sweep budget reasoning and emitted only the background glow, $1.60 for nothing. Given its own 128K ceiling on 2026-09-02 the same request finished in 74,061 tokens and $1.85 - its best laptop by a distance, and the first Opus 5 cell whose lid and deck belong to one object: silver body, notch, menu bar, a dock of real app icons, grilles, trackpad. The base is still a wedge that thins to nothing at the front, and the view is nearer flat-on than three-quarter. A clear step up from Opus 4.8's flat mid laptop in detail and effort-response - but what separates Fable is that it makes no visual errors, where Opus 5 only looks the part.

• high - All the detail is here - notch, dock with app icons, a code window, "MacBook Pro" under the screen. Read closely the construction does not hold up: the base is wider than the lid on both sides, sits at a far more top-down angle, and ends in a sharp point at the front-right. $0.50 and 20K output tokens, because thinking is on by default.
• xhigh - More detail and better materials than high, but the isometry is broken: the screen and the keyboard deck sit in different perspectives, so it looks slick at a glance and falls apart on a second look. 47K output tokens and $1.18 for it.
• max - At the model ceiling (128K): finished in 74,061 tokens and 14.9 minutes for $1.85. The richest Opus 5 drawing in content - silver body, notch, menu bar, a dock of real app icons, grilles, trackpad - and the most broken in construction. Every checklist run named the same five faults: sharp corners, a spike at a corner, a base that is a wedge rather than a slab, wrong proportions, and keys that never resolve into rows. Its high cell, at $0.50, is the one Opus 5 drawing no run faulted twice.
• max @64K - Cut off by OUR 64K sweep budget: thinking is on by default, so it spent almost all of it reasoning and emitted only the background glow, $1.60 billed for nothing. Given its own 128K ceiling it finished (above).

In Playcode since 2026-07-24 · runs 2026-07-24, 2026-09-02

The new challenger, and a serious one on finish: a proper dark MacBook with a dock, a notch and speaker grilles at both levels, for a fifth of Fable's price. On closer inspection it is not error-free, though. It draws the laptop nearly flat-on rather than in the three-quarter view asked for, the lid tapers while the base stays dead-on so the two perspectives disagree, and the base is a depthless slab whose ends splay past the lid. Tidy, cheap, and wrong in the way this test is built to expose - and the clock is brutal at max (17.5 minutes). As ever, the pelican it already aces at high gains nothing from max.

• high - A proper dark MacBook - dock, notch, keyboard, trackpad - already here for $0.15. That was a hair over Gemini 3.6 Flash at max until Google halved the Flash rate card; it is now roughly three times the dearest Gemini cell.
• max - The best-finished cheap laptop here - speaker grilles, a code window, a dock - but not error-free: it is drawn nearly flat-on instead of the three-quarter view asked for, and the tapered lid disagrees with the dead-on base. 30,579 tokens and 17.5 minutes for it.

In Playcode since 2026-07-23 · runs 2026-07-24

The first cheap model to draw a three-quarter MacBook that actually holds together: twilight wallpaper with a moon, a dock, a sculpted key grid, speaker grilles, and a hinge whose halves meet - the assembly GLM 5.2 and every Gemini never managed. The catch is the meter. That drawing took 73K output tokens, 16.5 minutes and $0.32 - forty-five times 5.2's price for the same task - because 5.3's server-side default effort sits between high and max, and reasoning bills as output. The dial, at least, is real: the first ORDERED effort ladder in this catalog, low 3.9K tokens to max 93K, the second-longest run on this page at 20 minutes. Its pelicans are the cheap tier's best, and max even gives the bird a saddle - the accessory nearly every other model forgets. Run it at low and it is a one-cent model that draws like a toy; leave the default and it spends like a mid-tier reasoner. Same $1.40/$4.40 card as 5.2.

• default - The best-assembled cheap laptop yet: a real three-quarter MacBook with a twilight wallpaper, dock, sculpted key grid and a hinge whose halves actually meet. It took 73K output tokens and 16.5 minutes - the server default effort sits between high and max.
• low - Crude at a cent and a half: the screen slab sits skewed on a flattened base.
• high - Front-on with a dated menu bar and full keyboard - and a ghost second lid edge behind the screen.
• max - 93K output tokens over 20 minutes - the second-longest run on this page. Tidier than high, same ghost slab.

In Playcode since 2026-09-04 · runs 2026-09-04

The surprise of the batch. Meta's Muse Spark 1.3, reachable only through OpenRouter, draws the second-best cheap laptop on the page: a menu bar with real Finder text, a notch, a browser window with traffic lights, a dock, a sculpted keyboard, a joined hinge - for three and a half cents. Its tell is a cyan glow spilling past the lid's edge and a slightly splayed deck. Its pain is latency, not price: up to 70 of a cell's seconds pass before the first token arrives through the aggregator, so a ninety-second drawing is mostly waiting. The pelicans are plainer flat cartoons - a helmet-and-cape rider at medium, no saddle anywhere - and the dial barely moves it: default, medium and high all land within 8.2-8.7K tokens on the laptop. Not selectable in Playcode; benchmarked because a frontier lab's cheap tier deserves a row.

• default - The surprise of the batch: menu bar with real Finder text, a notch, a browser window with traffic lights, a dock, a joined hinge - for three and a half cents. Its tells: a cyan glow spilling past the lid, and 59 of the 97 seconds passed before the first token.
• high - The dial barely moves it: default, medium and high all land within 8.2-8.7K tokens.

In Playcode since 2026-09-04 · runs 2026-09-04

Superseded by GLM 5.3, which finally adds the three-quarter view - at forty-five times the default-effort price. The most complete cheap laptop here, though not the best one. Everything is assembled correctly - notch, speaker grilles on both sides, a real key grid, a trackpad, a lid whose bezel meets the hinge - and nothing is detached or floating, which is more than Opus 5 at twenty times the price manages. Google has since halved the Flash rate card, so GLM's old price peer is gone: Gemini 3.6 Flash at low now costs $0.036 against GLM's $0.075, and 3.8 Flash packs in more detail at $0.049 - but every Gemini laptop is crooked somewhere, and GLM's is not. GLM still renders individual keys and grilles where Gemini smears the deck into a dark band, but it does not do the one thing both Geminis do - a proper three-quarter view. GLM ignores that instruction and draws the laptop nearly flat-on. Its base is skewed against the lid, sliding right as it comes forward, so the two halves do not share a vanishing point. The pelican inverts the failure: the bicycle is excellent - spoked wheels, a correct diamond frame, a crank - but the bird is draped over the top tube rather than sitting on it, with arrow-headed legs that stop short of the pedals. Same cost, same 17K tokens, same three minutes for both drawings: it has one budget and spends all of it whatever you ask for.

• default - Correctly assembled and cleanly shaded - notch, speaker grilles, key grid, trackpad - but drawn almost flat-on instead of the three-quarter view asked for, and the base is skewed against the lid. Only one cell: Z.ai has no effort dial, so the ladder does not apply.

In Playcode since 2026-06-19 · runs 2026-08-14

Better behaved than 3.7, and still not a well-built laptop. Every cell is misaligned somewhere: default's key block overhangs the deck's left edge and the two halves do not quite share a vanishing point, max's lid floats a hair off the hinge while 14.9K tokens - the heaviest Gemini Flash cell ever - go partly on ambient glow circles, and low smears the screen's colours across the keys. The parts are all present (screen glow, notch, menu bar, dock, a real key grid for $0.049) but the assembly never closes, which is the thing this page ranks. What 3.8 genuinely fixes is reliability: 8 cells, 8 valid SVGs, where 3.7 shipped a file no browser would finish - and the pelican keeps the tier's charm, cap, scarf, fish in the pouch, webbed foot on the crank. Google says 3.8 may think harder at higher efforts and the dial is still unordered: medium is its cheapest MacBook cell, max its dearest at nearly 2x. Same halved rate card as 3.7 through 2026-12-31. A crooked laptop for a fortieth of Fable's price is still a crooked laptop.

• default - Its best try: a three-quarter MacBook with screen glow, a notch, a menu bar, a dock and a real key grid - still misaligned, with the key block overhanging the deck edge and the two halves not quite sharing a vanishing point.
• low - The screen's colours smear across the keyboard deck - the weakest of its four cells, and dearer than default.
• medium - Its cheapest cell and arguably its cleanest: a tidy three-quarter laptop with an aligned key grid and speaker grilles.
• max - The heaviest Gemini Flash cell ever measured here - part of the budget goes on ambient glow circles behind the laptop, and the lid hovers a hair off its hinge line.

In Playcode since 2026-09-04 · runs 2026-09-04

Superseded by 3.8 Flash, which Google prices identically and which draws a straighter - still not straight - laptop. The most detailed draughtsman in the cheap tier by a distance, and the least reliable. Its MacBook has a notch camera, a menu bar, a dock of coloured icons and a real key grid for $0.052; its pelican genuinely rides the bicycle - helmet, scarf, a fish in the pouch, webbed feet on the pedals - which is more than most models here manage. Then it drops the assembly: the laptop's lid and base do not share a hinge, so the deck reads as a slab sliding out beneath the screen, and at max on the pelican it emitted invalid SVG that no browser will finish drawing. Read the price twice. It is a fortieth of Fable's best drawing, but ten to sixteen times Gemini 3 Flash for the same prompt, because it thinks eight times as hard and thinking bills as output. Effort buys nothing: its four cells scatter between 9.3K and 13.9K tokens in no order.

• max - The most detailed laptop in the benchmark - notch camera, menu bar, dock, a real key grid - but the lid and the base do not share a hinge, so the deck reads as a separate slab sliding out beneath it.

In Playcode since 2026-08-26 · runs 2026-08-26

The first DeepSeek that draws the machine. Where V4 Flash shipped orphan lines and gave up on perspective, every V4.1 Flash cell is a MacBook with a notch, screen content, speaker grilles and a trackpad, finished in 52 to 58 seconds - three times faster than its predecessor's laptops - for about two cents. None of the three holds together, and each fails differently: at low a huge grey slab hangs off the base like a second, empty laptop; at high the keyboard shears sideways off the deck and a black spike sticks out of the right edge; at max - the cleanest surface, with a menu bar, a dock and a macOS-style wallpaper - the lid is drawn from below and the base from above, with a translucent ghost of the deck floating off to the right. Assembly, not detail, is what keeps it out of the top ten. The dial barely moves the meter: 15.7K, 16.3K and 17.5K output tokens. Its pelicans are better than its laptops again, and low is the keeper. Released 2026-09-10 with native image input on a card cheaper than V4 Flash's ($0.30 in / $1.20 out).

• low - Two cents and 52 seconds for a recognisable MacBook - notch, a tiny wordmark on the chin, grilles both sides, a trackpad - drawn flat-on above a deck in steep perspective, with a huge grey slab hanging off the base like a second, empty laptop. The keyboard is horizontal bands, not keys.
• high - The most three-dimensional of the three, with individual keys, a dock and a screen glow on a dark ground - but the key block shears sideways and spills off the deck, the trackpad sits in a different projection from the keys, and a black spike sticks out of the right edge. Five seconds slower than low, for the same two cents.
• max - The cleanest finish DeepSeek has produced here: a menu bar with traffic lights, a dock, a macOS-style wallpaper, grilles, trackpad, a proper notch. Then the geometry: the lid is drawn from below and widens towards the top while the base is seen from above, the keyboard is a blank dark rectangle, and a translucent ghost of the deck floats off the right side. Its best laptop, and still a rank-12 laptop.

In Playcode since 2026-09-10 · runs 2026-09-10

The cheapest model in this catalog by a wide margin, and the first at this price to draw a machine you would recognise without being told. Its max cell is $0.019 - under two cents for 38,003 tokens and nine minutes - and it puts a notch and camera on the bezel, perforated speaker grilles on both sides, an individually-keyed keyboard and a properly proportioned trackpad on an aluminium body. What it does not do is furnish the screen: no menu bar, no dock, just a wallpaper, where models a few places above manage both. The assembly fault is consistent - the key block is drawn flatter than the deck it sits on, so the keyboard reads as pasted onto the laptop rather than set into it, with a thin wedge protruding past the front-left corner. Its dial is the sharpest illustration on this page of why token counts are not quality: output rises cleanly from 6.3K to 38K, but the DRAWINGS do not follow. Default, the cheapest cell at a third of a cent, has a full key grid. High, which costs nearly twice as much, has no keyboard at all - just a blank silver plate where the keys should be. Xhigh brings them back and adds grilles and a side port, but hangs the lid at a different eye height from the base. Only max gets everything into one object. It ranks above Grok 4.6 because its lid and base do at least join, which Grok's do not, and below DeepSeek V4.1 Flash, which furnishes its screens. Half the price of GPT-5.6 Luna, which it replaces on the Playcode Fast tier, and six places above it.

• default - A third of a cent. A recognisable laptop with a notch, a full key grid and a trackpad - but drawn nearly flat-on where the prompt asks for three-quarters, with no grilles and a wide empty stretch of deck to the right of the keyboard.
• high - Nearly twice the price of default and it has NO KEYBOARD - a blank silver plate where the keys should be, with only a faint rectangle marking the area and a single grille patch on the right. The clearest case on this page of more tokens buying a worse drawing.
• xhigh - Keys return, and with them dotted grilles on both sides and a port on the right edge. The lid is drawn flatter than the base, so the two halves sit at slightly different eye heights and the hinge does not quite close.
• max - Its best, and still under two cents: notch, grilles both sides, a full key grid, a trackpad and a bright wallpaper on an aluminium body. The key block is drawn flatter than the deck it sits in, and a thin white wedge juts past the front-left corner.

In Playcode since 2026-09-22 · runs 2026-09-23

The redemption of the xAI seat, partially. Grok 4.5 was this benchmark's floor - 668 output tokens for a laptop it barely attempted - and 4.6 actually engages: ten times the tokens, a screen with glow and a dock, a dense keyboard, a trackpad. What it cannot yet do is join the parts. At default the deck floats detached below the screen with a stray line poking from the left edge and the key lattice sliding off the deck's right side; at high the deck's corner tears away like a paper fold while the keyboard fades to near-white. The pelicans are tidier than the laptops: high is a calm, coherent flat cartoon done in 38 seconds for a cent - the fastest finished drawing on this page - and default puts two feet on the pedals of a solid dark-blue bike, though neither bird gets a saddle. Same $2/$6 card as 4.5, reasoning still always on, and the effort dial moves output in no consistent direction: high spent more tokens than default on the laptop and fewer on the bird.

• default - A real laptop at last from xAI - screen glow, dock, dense keyboard, trackpad - but the deck floats detached below the screen, the key lattice slides off its right edge, and a stray line pokes from the left.
• high - Calmer and paler: browser chrome and a camera notch on the screen, but the keyboard fades to near-white and the deck's right corner tears away like a paper fold.

In Playcode since 2026-09-04 · runs 2026-09-04

Superseded by 3.7 Flash, which Google prices identically. Two things changed since the launch sweep. Google halved the Flash rate card, so the run this article once billed at $0.16 is $0.041; and a fuller four-effort sweep killed the claim that this was the first Gemini with a working effort dial - its cells sit between 9.4K and 10.9K tokens in no order at all. What survives is the drawing: flat-shaded, plainer than 3.7, and correct everywhere it commits, though it smears the MacBook keyboard into dark bands. It broke once, emitting invalid SVG on the MacBook at default. It still takes the pelican off 3.7, which broke on that one.

• default - Emitted invalid SVG - the error was: Specification mandates value for attribute points. A browser draws only as far as the error, so there is no finished picture.

In Playcode since 2026-07-21 · runs 2026-08-26

Superseded by V4.1 Flash, which DeepSeek prices lower and which finally draws the machine - see rank 12. The single clearest piece of evidence that the pelican test is contaminated. Its pelican at max is genuinely good - a properly built bicycle with a chain and cranks, and a pelican actually sitting on it - and its MacBooks are among the worst drawings in this entire benchmark. Not merely weak: broken. Stray lines and orphan shapes drift across the canvas, the lid and the keyboard deck are drawn in different perspectives, key rows shear off to one side. Even its best laptop gives up on the three-quarter view the prompt asks for and retreats to a flat front-on cartoon with a keyboard too narrow for its own deck. That gap is not a fluke of effort - a model this cheap and this fluent at the bird cannot draw the object nobody has trained it on, which is the whole reason we stopped grading on pelicans. Costs about a cent per drawing, and on the MacBook that is still not a bargain.

• low - Three quarters of a cent, and the least broken of its three - which is the only sense in which it is good. A generic silver wedge: the lid is a paper-thin sliver with no depth at all, the screen dwarfs the body, and nothing marks it as a MacBook rather than any laptop in a stock-vector pack.
• high - One of the worst drawings in this benchmark, at any price. The lid and the keyboard deck are in two different perspectives, the key rows shear away to the right and spill past the deck, the trackpad is a skewed parallelogram, and loose lines and orphan shapes hang off the front edge and the hinge. It is also the cheapest of its three runs - raising the dial one notch from low makes the picture visibly worse.
• max - Double the tokens and double the wait to stop drawing stray lines. It finally names the machine - notch, "MacBook Pro" wordmark, symmetric body - but only by abandoning the brief: the prompt asks for a 3D model and this is a flat, front-on cartoon with no perspective at all, its keyboard a narrow trapezoid floating on a deck it does not fit. The best MacBook it can draw, and still near the bottom of the field.

In Playcode since 2026-05-27 · runs 2026-08-02

Cheap, fast, consistent - and consistently mid. Its three effort levels cost $0.05, $0.10 and $0.18 and produce what is recognisably the same drawing. The effort dial buys nothing here.

In Playcode since 2026-05-31 · runs 2026-07-13

The cheap-and-fast control, and still the bargain of the benchmark: a simplified laptop with flat shading where nothing is broken, for $0.005. It knows what it can execute and stays inside that envelope - eight cells across both tasks and not one line of invalid markup, which neither Gemini heavyweight manages. The effort dial does nothing to it: max is its cheapest cell of the four, not its dearest.

• max - The dial is inert on this model: max is its CHEAPEST cell of the four, not its dearest.

In Playcode since 2026-01-20 · runs 2026-08-26

Superseded by GPT-6 Luna at half the price again ($0.20/$1.20 -> $0.10/$0.50), which draws a better laptop six places above - see rank 15. The cheapest OpenAI entry ever on this page until then. After the 2026-09-04 reprice ($1/$6 cut to $0.20/$1.20) it competes with Gemini 3 Flash on price - its default MacBook costs half a cent - and what that buys is uneven. xhigh draws its best laptop, a rounded silver body with a sculpted keyboard, coherent and assembled if a little toy-ish; high smears rectangles of noise texture across the deck; max mistakes the brief for a tablet propped on a serving tray. Like GLM 5.3 the same day, its dial is genuinely ordered - 4.2K tokens at default rising monotonically to 23.4K at max, on both tasks - which no OpenAI sibling here manages. The claim that it matches Gemini 3 Flash is close but not true: Flash never breaks, Luna sometimes does. It ranks one below.

• default - A recognisable laptop for half a cent on the freshly cut nano pricing - the base is a thin angular wedge and the keyboard floats on it.
• high - Ships rectangles of noise-texture static across the deck. Valid SVG, broken picture.
• xhigh - Its best cell: a rounded silver body with a sculpted keyboard - coherent, assembled, a little toy-ish.
• max - Mistakes the brief for a tablet propped on a serving tray.

In Playcode since 2026-09-04 · runs 2026-09-04

Rejects xhigh outright - the bug this benchmark found in its first ninety seconds. What it does draw is recognisable but loose, with parts in the wrong places under inspection.

In Playcode since 2026-06-29 · runs 2026-07-13

Also rejects xhigh. Draws a laptop-shaped object with misplaced parts, and effort barely moves it: $0.11, $0.11, $0.10 across the ladder.

In Playcode since 2026-03-02 · runs 2026-07-13

Fast and reliable, and it does not know what a MacBook looks like. Even uncapped at max, with 29,834 tokens to spend, it draws a wide, thin generic ultrabook. The clearest evidence in this benchmark that budget does not buy understanding.

• max - Uncapped it finishes - and the form is still wrong. Budget does not buy understanding.

In Playcode since 2026-07-09 · runs 2026-07-13, 2026-07-21

Broken geometry on the laptop, and the most expensive way to get there in tokens: at max it needs 121,713 output tokens and 21 minutes to close one SVG - nearly double our old 64K sweep budget, which is why this cell read as a runaway until 2026-09-02. What it finally draws is a complete laptop, dock and key grid included, tipped over: the lid and the base are two skewed slabs sharing no hinge and no vanishing point. Perfectly fine on a pelican, lost on an object you know well.

• max - At the model ceiling (128K): it closes the SVG at last - 121,713 tokens, 21 minutes, $1.22, the longest run on this page. A complete laptop with a dock, a window and a full key grid, drawn tipped over: the lid and the base are two skewed slabs that share no hinge and no vanishing point.
• max @64K - Cut off by OUR ~64K sweep budget (the provider reported 65,538 output tokens against it): it never closed the SVG. At its own 128K ceiling it needed 121,713 tokens - nearly double our budget.

In Playcode since 2026-06-30 · runs 2026-07-13, 2026-09-02, 2026-07-21

Superseded by Grok 4.6, which xAI prices identically and which actually tries. Barely tries. 668 output tokens on the MacBook - more logo than laptop - though that makes it the cheapest run in the benchmark at $0.004. Its pelican is fine. high is its ceiling.

• high - high is its ceiling. 668 output tokens - it barely tried.

In Playcode since 2026-07-10 · runs 2026-07-13