2026-07-11 3d

The Big Mac test: twelve models, one Big Mac

One prompt, no edits: whatever came back is published — live sandboxed artifacts (drag to rotate), metrics and prices.

The prompt

promptCreate a 3D visualization of a Big Mac in three.js. The burger slowly rotates on its own; the camera can orbit and zoom. Make it as photorealistic as possible.

Technical requirements: a single self-contained HTML file; load three.js from a CDN; the scene fills the entire browser window and adapts to its size; no localStorage, cookies, or network requests beyond CDN libraries. Return only the finished HTML file as a single code block, with no text before or after.

Results

Score is a weighted blend: 0.4 big-mac-ness, 0.3 render quality, 0.2 interactivity, 0.1 first-run success. Points are awarded by place in this test using the 10, 7, 5, 3, 2, 1 scheme; lower places score nothing. A tie does not split a place: the cheaper run goes higher, then the faster one, and the last key is the snapshot id alphabetically.

#	model	score	pts	cost	value	time
01	claude-opus-5🏆	0.875	+10	$1.31	0.7	617s
02	gpt-5.6-sol	0.840	+7	$0.36	2.3	141s
03	qwen3.8-max	0.830	+5	$0.22	3.8	751s
04	gpt-5.5	0.800	+3	$0.63	1.3	238s
05	claude-fable-5	0.745	+2	$0.98	0.8	234s
06	deepseek-v4-pro	0.740	+1	$0.03	25.9	403s
07	kimi-k3	0.705	—	$0.53	1.3	1162s
08	gemini-3.1-pro	0.640	—	$0.16	3.9	99s
09	claude-opus-4-8	0.635	—	$0.15	4.3	57s
10	grok-4.5	0.615	—	$0.07	8.7	68s
11	claude-opus-4-6	0.530	—	$1.49	0.4	864s
12	qwen3.7-maxdidn't render	0.000	—	$0.03	0.0	149s
	total			$5.96		79:43

value = score ÷ cost (single run) — a rough “score per dollar.” score is half-subjective and cost varies between runs, so treat it as a hint, not a hard metric. The gap is real, though: deepseek lands 0.14 behind the winner at ~a fortieth of the cost.

Compare two

Defaults to the winner vs the best value; pick any pair from the menus.

Artifacts

01 claude-opus-5 winner 0.875

tokens	cost	time	code	file
160 → 52.4k	$1.31	617s	806 loc	32 KB

The full spec, finally: a three-part bun with real crust mottling and properly seated sesame, two grained patties, cheese with actual thickness, shredded lettuce, onion, pickles. Best materials of the twelve — and the only entry whose garnish explodes: lettuce shards and cheese wings stick out a third past the bun. 10 minutes and $1.31, second-priciest of the run. Two disclosures: it was delivered on the second attempt (the first stream was cut at ~100s by the transport, not by the model — one of the run's two reruns), and its two subjective marks were set on the owner's behalf rather than by his own eye.

open ↗ · raw response

02 gpt-5.6-sol 0.840

tokens	cost	time	code	file
108 → 12.0k	$0.36	141s	739 loc	25 KB

The clearest Big Mac of the first ten: two patties, a proper middle bun, white-sauce specks — and the most convincing light of them. Won the first ten on structure, not just polish.

open ↗ · raw response

03 qwen3.8-max 0.830

tokens	cost	time	code	file
149 → 36.8k	$0.22	751s	509 loc	22 KB

One version on from the sibling that failed: qwen3.7-max pulled dead CDN paths, qwen3.8-max ships an import map for three@0.165.0 and ESM addons — and it runs. Third place. Big-mac-ness level with claude-opus-5 (0.80): three-part bun, two patties, lettuce on both levels, honest big sesame, onion crumb and sauce dots — the cheese is even modelled with a melted droop, but it is buried under lettuce sheets grown too wide, and pickles are the one thing missing. The render (0.70) is plasticine rather than meat: near-black patties, mint-green lettuce, sesame in blobs. $0.22 — cheaper than everything above it. 12.5 minutes, third-slowest of the twelve.

open ↗ · raw response

04 gpt-5.5 0.800

tokens	cost	time	code	file
108 → 21.1k	$0.63	238s	707 loc	25 KB

Same double-decker-plus-sauce reading as gpt-5.6-sol, a step behind on lighting and materials. Solid, unspectacular.

open ↗ · raw response

05 claude-fable-5 0.745

tokens	cost	time	code	file
160 → 19.5k	$0.98	234s	364 loc	15 KB

Gets the two-patty, middle-bun structure right — the lettuce is flat, stylized sheets rather than leaves. $0.98 for fifth, pricier than both GPTs above it.

open ↗ · raw response

06 deepseek-v4-pro 0.740

tokens	cost	time	code	file
106 → 22.3k	$0.03	403s	793 loc	38 KB

Mid-table (0.740) for $0.03 — about a tenth of gpt-5.6-sol's $0.36. A double-decker on a plate, fillings a touch sparse. The cheapest strong result of the run.

open ↗ · raw response

07 kimi-k3 0.705

tokens	cost	time	code	file
186 → 35.1k	$0.53	1162s	475 loc	20 KB

The slowest by far: 19 minutes of reasoning before a single line of HTML — it timed out on the first run at the standard 15-minute budget and only finished on a doubled one (one of the run's two reruns). Nails the Big Mac structure — two patties, a middle bun, sesame top — but dark wood-like patties and sparse lettuce keep it mid-table. $0.53. The one entry with a doubled time budget; its two subjective marks were also set on the owner's behalf rather than by his own eye.

open ↗ · raw response

08 gemini-3.1-pro 0.640

tokens	cost	time	code	file
102 → 13.5k	$0.16	99s	478 loc	23 KB

Clean single-patty burger, but rendered small in the frame — asked to fill the window, it didn't quite. Fastest of the mid-pack at 99s.

open ↗ · raw response

09 claude-opus-4-8 0.635

tokens	cost	time	code	file
160 → 5.9k	$0.15	57s	368 loc	12 KB

The best bun of the first ten — real sesame texture, soft shading — around a single-patty cheeseburger, not a Big Mac. Craft high, big-mac-ness low. Also the fastest: 57s.

open ↗ · raw response

10 grok-4.5 0.615

tokens	cost	time	code	file
308 → 11.7k	$0.07	68s	328 loc	9 KB

A double-decker, but geometric and bare — no lettuce, no sauce. Second-cheapest of the renderers at $0.07, behind deepseek-v4-pro's $0.03.

open ↗ · raw response

11 claude-opus-4-6 0.530

tokens	cost	time	code	file
114 → 59.6k	$1.49	864s	334 loc	14 KB

The cautionary tale: 14 minutes and 59.6k tokens — the most expensive run ($1.49) — for a plain domed bun with no visible patties. Effort ≠ result.

open ↗ · raw response

12 qwen3.7-max 0.000

didn't render
THREE is not defined

tokens	cost	time	code	file
111 → 8.2k	$0.03	149s	660 loc	26 KB

Didn't render — THREE is not defined. It pulled three.js from dead CDN paths (the old global three.min.js and examples/js/OrbitControls.js, both 404 since three.js dropped them around r150). A real signal: stale knowledge of how three.js ships today.

open ↗ · raw response

run conditions and versions

12/12 responded · 299.8k tok · $5.96 · 84:42 wall

one-shot · single-delivery · temperature default · reasoning high · max_tokens uncapped · 2026-07-11

how score is computed. The owner sets big-mac-ness and render quality by eye, on a 0–1 rubric. A headless-browser probe produces first-run success and interactivity (loaded without errors + responds to drag). The subjective part is disclosed honestly: it is an author’s judgment, not an “objective” number — so the leaderboard order and the value column inherit that subjectivity.

re-run by the owner: the first attempt produced no working artifact (no usable response / truncated / broken), this result is a re-attempt, single-delivery held per attempt: kimi-k3, claude-opus-5

snapshots: claude-opus-4-8=anthropic/claude-opus-4.8 · claude-opus-4-6=anthropic/claude-opus-4.6 · claude-fable-5=anthropic/claude-fable-5 · gpt-5.5=openai/gpt-5.5 · gpt-5.6-sol=openai/gpt-5.6-sol · gemini-3.1-pro=google/gemini-3.1-pro-preview · grok-4.5=x-ai/grok-4.5 · qwen3.7-max=qwen/qwen3.7-max · deepseek-v4-pro=deepseek/deepseek-v4-pro · kimi-k3=moonshotai/kimi-k3 · claude-opus-5=anthropic/claude-opus-5 · qwen3.8-max=qwen/qwen3.8-max

← all tests