promptsAI Lab

Which AI model drew the best map of New York? Results of our test

Results of the same prompt side by side. Nothing runs on this page – click a row to open the output. Sort by clicking a column header; the best value is highlighted.

Short answer

Tested 14 Oct 2026 – 6 Nov 2026, one attempt for each of the 40 outputs, model versions are in the table. This is the result of our test, not a general measure of a model's reliability – another attempt may turn out differently. Time = from sending the prompt to the last token of the answer, thinking included (for runs in Claude Code the whole task incl. writing the file), fixes excluded. Cost = tokens × the public API price list as of the date at the model, fixes excluded; ≈ = an estimate when only the token total is known. It is not the amount actually paid – subscriptions, caching or discounts change it.

⇆ Live side by side
Grok 4.7 default
xAI
Model profile →
loads958/89987 / 8051 min 52 s243,331$1.45
prices as of 24/10/2026
161 kB
2,919 lines
Gemini 4 Argon medium
Google
Model profile →
loads938/89884 / 706 min 34 s40,036$0.40
prices as of 17/10/2026
103 kB
2,825 lines
Claude Sonnet 5.5 medium
Anthropic
Model profile →
loads918/89880 / 6037 min 03 s228,876≈ $2.11
prices as of 06/10/2026
98 kB
1,143 lines
GPT-6 Astra medium
OpenAI
Model profile →
loads918/89562 / 809 min 17 s229,263$1.50
prices as of 15/10/2026
57 kB
233 lines
GPT-6.1 Sol medium
OpenAI
Model profile →
loads887.5/89467 / 806 min 54 s191,285$0.21
prices as of 15/10/2026
42 kB
130 lines
GPT-5.5 medium
OpenAI
Model profile →
loads887.5/89873 / 703 min 02 s18,651$0.55
prices as of 17/10/2026
42 kB
699 lines
Claude Haiku 5.5 medium
Anthropic
Model profile →
loads877.5/89878 / 609 min 52 s199,031≈ $0.090
prices as of 06/10/2026
41 kB
653 lines
DeepSeek V4.1 Flash 1017 high
DeepSeek
Model profile →
loads837.5/89835 / 809 min 30 s134,049$0.080
prices as of 20/10/2026
107 kB
2,490 lines
Kimi K3.1 Pro thinking
Moonshot AI
Model profile →
loads827/89467 / 7032 min 09 s74,126$0.74
prices as of 25/10/2026
101 kB
1,502 lines
GPT-6 Sol medium
OpenAI
Model profile →
loads806.5/810070 / 703 min 13 s227,541$0.24
prices as of 15/10/2026
37 kB
189 lines
GPT-5.6 Sol medium
OpenAI
Model profile →
loads796.5/89462 / 802 min 46 s203,462$0.34
prices as of 15/10/2026
29 kB
102 lines
Grok 4.6 default
xAI
Model profile →
loads796.5/88167 / 8018 min 21 s72,150$0.43
prices as of 18/10/2026
76 kB
1,386 lines
Muse Wonder thinking
Meta
Model profile →
loads796.5/89467 / 702 min 13 s19,583$0.38
prices as of 25/10/2026
35 kB
253 lines
Gemini 3.8 Flash medium
Google
Model profile →
loads776.5/88473 / 604 min 03 s31,029$0.12
prices as of 16/10/2026
67 kB
1,762 lines
Gemini 3.6 Flash medium
Google
Model profile →
loads766/89889 / 502 min 17 s26,897$0.100
prices as of 17/10/2026
65 kB
1,632 lines
GLM-5.3-Flash high
Z.ai
Model profile →
loads766/89467 / 808 min 22 s107,863$0.054
prices as of 31/10/2026
70 kB
1,025 lines
Gemini 3.1 Pro medium
Google
Model profile →
loads746/89681 / 501 min 55 s17,987$0.21
prices as of 17/10/2026
32 kB
745 lines
Qwen3.6 27B thinking
Alibaba
Model profile →
loads705.5/89879 / 505 min 35 s24,889$0.049
prices as of 25/10/2026
63 kB
1,467 lines
Hy4 preview thinking
Tencent
Model profile →
loads706/89871 / 309 min 05 s40,803$0.10
prices as of 06/11/2026
26 kB
467 lines
GPT-5.4 mini medium
OpenAI
Model profile →
loads666/89567 / 107 min 15 s118,367$0.53
prices as of 17/10/2026
94 kB
2,250 lines
Gemma 4 31B default
Google
Model profile →
loads624.5/89883 / 404 min 09 s9,177$0.00
prices as of 17/10/2026
23 kB
534 lines
DeepSeek V4.1 Flash 0910 high
DeepSeek
Model profile →
loads614.5/89887 / 306 min 57 s93,586$0.056
prices as of 16/10/2026
82 kB
1,701 lines
MiniMax M3.1 Flash thinking
MiniMax
Model profile →
loads594.5/89359 / 507 min 34 s57,588$0.034
prices as of 25/10/2026
94 kB
2,057 lines
Gemini 3.5 Flash-Lite medium
Google
Model profile →
loads584.5/89786 / 1048 s16,156$0.039
prices as of 17/10/2026
47 kB
1,231 lines
GPT-6 Luna medium
OpenAI
Model profile →
loads574/89862 / 501 min 22 s228,992$0.00800
prices as of 15/10/2026
25 kB
65 lines
Mistral Medium 3.5 high
Mistral
Model profile →
loads514/89167 / 105 min 32 s51,330$0.38
prices as of 17/10/2026
43 kB
1,190 lines
Mistral Small 4 high
Mistral
Model profile →
loads503.5/88386 / 202 min 02 s24,465$0.014
prices as of 17/10/2026
48 kB
1,412 lines
DeepSeek V4 Pro high
DeepSeek
Model profile →
loads443/88276 / 103 min 24 s31,746$0.062
prices as of 16/10/2026
59 kB
1,566 lines
Codestral 25.08 default
Mistral
Model profile →
loads443/810078 / 01 min 14 s8,234$0.00714
prices as of 17/10/2026
31 kB
688 lines
Gemini 3.1 Flash-Lite medium
Google
Model profile →
loads412.5/89250 / 409 s3,996$0.00544
prices as of 17/10/2026
9 kB
184 lines
Claude Opus 5.5 medium
Anthropic
Model profile →
broken
fixed version →
0
after fix 95
0/8
after fix 8/8
70
after fix 99
0 / 0
after fix 79 / 90
49 min 23 s235,314≈ $4.33
prices as of 06/10/2026
119 kB
1,703 lines
GPT-5.6 Terra medium
OpenAI
Model profile →
broken
fixed version →
0
after fix 71
0/8
after fix 5.5/8
71
after fix 100
0 / 0
after fix 69 / 60
1 min 26 s10,564$0.12
prices as of 17/10/2026
27 kB
357 lines
GPT-5.6 Luna medium
OpenAI
Model profile →
broken
fixed version →
0
after fix 67
0/8
after fix 5.5/8
55
after fix 80
0 / 0
after fix 67 / 50
48 s6,880$0.00782
prices as of 17/10/2026
18 kB
159 lines
GLM 5.3 high
Z.ai
Model profile →
broken00/8830 / 09 min 56 s128,440$0.56
prices as of 17/10/2026
30 kB
444 lines
Mistral Large 4 Preview none
Mistral
Model profile →
broken00/8–– / 024 min 22 s128,449$0.27
prices as of 17/10/2026
–
Kimi K3 thinking
Moonshot AI
Model profile →
broken
fixed version →
0
after fix 84
0/8
after fix 7/8
85
after fix 100
81 / 0
after fix 80 / 60
49 min 07 s88,683$1.32
prices as of 22/10/2026
82 kB
1,002 lines
Qwen3.8 Max thinking
Alibaba
Model profile →
broken
fixed version →
0
after fix 83
0/8
after fix 7/8
71
after fix 99
0 / 0
after fix 59 / 80
35 min 39 s96,890$0.58
prices as of 22/10/2026
75 kB
975 lines
MiMo-V2.6-Pro thinking
Xiaomi
Model profile →
broken
fixed version →
0
after fix 3
0/8
after fix 0/8
–
after fix 29
– / 0
after fix 0 / 0
13 min 34 s102,520$0.089
prices as of 25/10/2026
81 kB
1,455 lines
GPT-6.1 Terra medium
OpenAI
Model profile →
broken
fixed version →
0
after fix 80
0/8
after fix 6.5/8
69
after fix 94
0 / 0
after fix 67 / 80
3 min 12 s17,688$0.087
prices as of 26/10/2026
45 kB
524 lines
Mistral Large 4 1029 medium
Mistral
Model profile →
broken
fixed version →
0
after fix 53
0/8
after fix 3.5/8
96
after fix 96
81 / 0
after fix 81 / 30
6 min 32 s58,492$0.12
prices as of 31/10/2026
36 kB
595 lines

Cost = tokens priced with the API price list valid on the date shown next to the cost (prices change). ≈ = estimate (only the total token count is known), no ≈ = exact from input / output tokens. “Loads” = the page opens without errors; technical health also checks behaviour during use – judge the quality in the viewer.

Evaluation

40 outputs from 40 models for the same zero-shot prompt – each one the model's first answer, published unedited.

  • 4.4/8average score
  • 30/40working outputs
  • 90/100avg. technical health
  • 61/100avg. UX

Broken (do not run): 10, 8 of them with a copy fixed by the same model. The average checklist score is 4.4/8 (a broken output scores 0), 5.8/8 among working ones. Best result in our test: Claude Sonnet 5.5, GPT-6 Astra, Gemini 4 Argon and 1 more – 8/8. Weakest working: Gemini 3.1 Flash-Lite – 2.5/8.

How we evaluate – methodology →

How the 8 requirements were met

From 30 working outputs: ✓ met, ½ partly, ✗ missing. Rate = (✓ + ½ × 0.5) / number of outputs.

Requirement✓½✗Rate
Recognizable Manhattan, rivers, neighbouring areas, Roosevelt Island, bridges 7176 52 %
Rotated street grid, avenues, major cross streets, Broadway, irregular Lower Manhattan 8148 50 %
Detailed Central Park (Reservoir, lakes, lawns, Ramble, Mall, Bethesda, Belvedere, Zoo, paths) 15114 68 %
30+ landmarks roughly in place, categories with icons 19110 82 %
Pan & zoom with mouse, wheel, pinch, buttons, keyboard; zoom-dependent labels 19110 82 %
Hover tooltip, info card on click, search with suggestions and fly-to 18120 80 %
Layer toggles, scale bar, compass, guided tour 17130 78 %
Drawn in code, full window, resize & high-DPI, runs without errors 2550 92 %

Strengths

The requirements the models met best – quotes from the reviewer's verdicts:

Drawn in code, full window, resize & high-DPI, runs without errors · 92 % met

  • Code-drawn, full window, DPR-aware resize, renders fine. – Claude Sonnet 5.5
  • Code-drawn SVG, full window, resize handled, crisp high-DPI. – GPT-6 Astra

30+ landmarks roughly in place, categories with icons · 82 % met

  • Over 30 correctly placed landmarks with category icons. – Claude Sonnet 5.5
  • 55 landmarks at real coordinates with category icons. – GPT-6 Astra

Typical mistakes

The requirements the models most often missed or met only partly:

Rotated street grid, avenues, major cross streets, Broadway, irregular Lower Manhattan · 50 % met

  • Streets have no stroke style; invisible on the map. – DeepSeek V4.1 Flash 0910
  • Streets render as black wedges; no visible grid. – MiniMax M3.1 Flash

Recognizable Manhattan, rivers, neighbouring areas, Roosevelt Island, bridges · 52 % met

  • Manhattan is an extremely thin sliver (1:37), the neighbours are grey slabs. – Hy4 preview
  • Crude polygon, rectangle neighbours, and the map is upside down. – Mistral Medium 3.5

Detailed Central Park (Reservoir, lakes, lawns, Ramble, Mall, Bethesda, Belvedere, Zoo, paths) · 68 % met

  • Detailed park data renders as an undifferentiated black area. – DeepSeek V4.1 Flash 0910
  • Park features defined but render invisibly tiny in corner. – DeepSeek V4 Pro

Agreement of two independent reviewers

The checklist of 40 outputs was filled in independently by two AI reviewers, Claude Opus 5.5 and Codex (GPT-6 Astra, medium). They agreed on 250 of 320 items (78 %). Average score by Claude 5.6/8, by Codex 5.2/8. Codex gave a lower score for 21, a higher one for 10 and the same for 9 outputs. They disagreed most often on: Recognizable Manhattan, rivers, neighbouring areas, Roosevelt Island, bridges (11×), Pan & zoom with mouse, wheel, pinch, buttons, keyboard; zoom-dependent labels (11×), Detailed Central Park (Reservoir, lakes, lawns, Ramble, Mall, Bethesda, Belvedere, Zoo, paths) (10×).

Where the reviewers disagree, the stricter (lower) verdict is proposed; the published score is approved by a human.

Objective tests

Measured by a tool, not by a person or an AI – so they are shown even before the scores are approved.

  • Tech: average 90/100 from 38 measurements; highest GPT-6 Sol (100), lowest GPT-5.6 Luna (55). Measures Lighthouse, smoothness, errors during automated use, the mobile layout and whether the controls work.
  • UX / UI: average 61/100 from 38 measurements; highest Gemini 3.6 Flash (89), lowest GPT-6.1 Terra (0). Measures control size, layout at 3 widths, keyboard, feedback and readability.