Rams Index

How well frontier models design interfaces. Every model builds the same pages from the same briefs, and Rams grades each rendered page at desktop and phone size. It is our opinion, held to a fixed method and checked against designers.

Models
6
Pages built
75
Judge passes
90
Briefs
5 + 0 private
Viewports
1440 · 390 px
Engine
f0bb659
Designer check pending · read Nov 6Scored by Rams eyes · Claude Sonnet 4.6Cross-judge GPT-5.6 TerraTrack: Rams Bench v0 · public suite v2

Top five

Rams Index, 0 to 100, with the 95% interval. Cost is generation only, per page.

  1. Rank 1=OpenAI

    GPT-6.1 Sol

    65.4
    55.7–73.9$0.15 per page
  2. Rank 1=Anthropic

    Claude Opus 5.5

    60.4
    52.9–70.2$0.28 per page
  3. Rank 1=OpenAI

    GPT-5.6 Sol

    58.9
    51.6–66.9$0.17 per page
  4. Rank 2=Anthropic

    Claude Fable 5.1

    58.9
    52.0–65.9$0.87 per page
  5. Rank 2=Anthropic

    Claude Fable 5

    57.6
    51.0–67.0$0.53 per page

Leaderboard

Rank is 1 plus the number of models clearly better (paired 95% interval above zero). Overlapping models share a rank.

Sorted by rank, ascending
  1. 1=GPT-6.1 SolOpenAI · 2026-0965.455.7–73.9

    Provisional: 1.0 judge passes per page, protocol 2

    Cost and checks

    $ per page
    $0.15
    Grade sum, 7–35
    25.30
    Terra, cross-judge
    79.8
    Visual, secondary
    88.8
    Defects per page
    5.2
    Pages · passes
    10 · 10

    Index by brief

    Landing
    70.5
    Dashboard
    68.8
    Settings
    47.3
    Pricing
    61.6
    Article
    78.6

    Grades by dimension, 1–5

    Hierarchy
    3.80
    Type
    3.65
    Spacing
    3.25
    Color
    3.70
    Components
    3.60
    Composure
    3.60
    Craft
    3.70
  2. 1=Claude Opus 5.5Anthropic · 2026-0960.452.9–70.2

    Provisional: 1.0 judge passes per page, protocol 2

    Cost and checks

    $ per page
    $0.28
    Grade sum, 7–35
    23.90
    Terra, cross-judge
    70.7
    Visual, secondary
    86.5
    Defects per page
    3.2
    Pages · passes
    10 · 10

    Index by brief

    Landing
    62.5
    Dashboard
    58.9
    Settings
    51.8
    Pricing
    50.9
    Article
    77.7

    Grades by dimension, 1–5

    Hierarchy
    3.60
    Type
    3.40
    Spacing
    3.15
    Color
    3.50
    Components
    3.40
    Composure
    3.45
    Craft
    3.40
  3. 1=GPT-5.6 SolOpenAI · 2026-0658.951.6–66.9

    Provisional: 1.3 judge passes per page, protocol 2

    Cost and checks

    $ per page
    $0.17
    Grade sum, 7–35
    23.50
    Terra, cross-judge
    79.3
    Visual, secondary
    88.6
    Defects per page
    4.7
    Pages · passes
    15 · 20

    Index by brief

    Landing
    67.0
    Dashboard
    52.1
    Settings
    52.7
    Pricing
    51.2
    Article
    71.7

    Grades by dimension, 1–5

    Hierarchy
    3.40
    Type
    3.20
    Spacing
    3.18
    Color
    3.33
    Components
    3.57
    Composure
    3.43
    Craft
    3.38
  4. 2=Claude Fable 5.1Anthropic · 2026-0858.952.0–65.9

    Provisional: 1.0 judge passes per page, protocol 2

    Cost and checks

    $ per page
    $0.87
    Grade sum, 7–35
    23.50
    Terra, cross-judge
    77.1
    Visual, secondary
    88.5
    Defects per page
    3.8
    Pages · passes
    10 · 10

    Index by brief

    Landing
    67.9
    Dashboard
    58.0
    Settings
    50.9
    Pricing
    50.0
    Article
    67.9

    Grades by dimension, 1–5

    Hierarchy
    3.45
    Type
    3.35
    Spacing
    3.10
    Color
    3.35
    Components
    3.35
    Composure
    3.50
    Craft
    3.40
  5. 2=Claude Fable 5Anthropic · 2026-0657.651.0–67.0

    Provisional: 1.3 judge passes per page, protocol 2

    Cost and checks

    $ per page
    $0.53
    Grade sum, 7–35
    23.13
    Terra, cross-judge
    74.5
    Visual, secondary
    88.7
    Defects per page
    3.2
    Pages · passes
    15 · 20

    Index by brief

    Landing
    59.5
    Dashboard
    56.0
    Settings
    49.4
    Pricing
    49.7
    Article
    73.5

    Grades by dimension, 1–5

    Hierarchy
    3.33
    Type
    3.28
    Spacing
    3.20
    Color
    3.27
    Components
    3.35
    Composure
    3.38
    Craft
    3.32
  6. 6Claude Opus 4.8Anthropic · 2026-0552.948.0–60.5

    Provisional: 1.3 judge passes per page, protocol 2

    Cost and checks

    $ per page
    $0.15
    Grade sum, 7–35
    21.80
    Terra, cross-judge
    69.5
    Visual, secondary
    85.2
    Defects per page
    2.5
    Pages · passes
    15 · 20

    Index by brief

    Landing
    51.5
    Dashboard
    48.8
    Settings
    47.3
    Pricing
    49.1
    Article
    67.6

    Grades by dimension, 1–5

    Hierarchy
    3.25
    Type
    3.10
    Spacing
    3.05
    Color
    3.05
    Components
    3.02
    Composure
    3.17
    Craft
    3.17
0–10045–75Generation7–35IndexIndexIndexIndexIndexCross-judgeSecondaryPer pagePasses
1=GPT-6.1 SolOpenAI · 2026-0965.4
55.7–73.9
$0.1525.3070.568.847.361.678.679.888.85.210 · 10
1=Claude Opus 5.5Anthropic · 2026-0960.4
52.9–70.2
$0.2823.9062.558.951.850.977.770.786.53.210 · 10
1=GPT-5.6 SolOpenAI · 2026-0658.9
51.6–66.9
$0.1723.5067.052.152.751.271.779.388.64.715 · 20
2=Claude Fable 5.1Anthropic · 2026-0858.9
52.0–65.9
$0.8723.5067.958.050.950.067.977.188.53.810 · 10
2=Claude Fable 5Anthropic · 2026-0657.6
51.0–67.0
$0.5323.1359.556.049.449.773.574.588.73.215 · 20
6Claude Opus 4.8Anthropic · 2026-0552.9
48.0–60.5
$0.1521.8051.548.847.349.167.669.585.22.515 · 20
  • Rams Index = (grade sum − 7) / 28 × 100. The grade sum is seven 1–5 grades from Rams eyes · Claude Sonnet 4.6, averaged over desktop and phone, then passes, then briefs.
  • * Provisional, every row in v0: fewer judge passes per page, protocol 2 (v0 ran 1.0–1.3). Tap a star for the model's reason. A starred rank is never quoted.
  • Terra is the same pipeline with GPT-5.6 Terra as the judge, so a reader can check whether a Claude judge flatters Claude pages. Visual is the judge's finding-based score, secondary.
  • Harness: one single-file HTML page per build, the same brief text for every model, provider default reasoning settings, every build judged as rendered.

Frontier

Rams Index against generation cost per page, log scale.

Scatter of Rams Index against cost per page for each model; filled points are on the frontier. The list beside it carries the same values.45505560657075$0.1$0.3$1$ per page (generation)Rams IndexGPT-6.1 Sol · 65.4 · $0.15Claude Opus 5.5 · 60.4 · $0.28GPT-5.6 Sol · 58.9 · $0.17Claude Fable 5.1 · 58.9 · $0.87Claude Fable 5 · 57.6 · $0.53Claude Opus 4.8 · 52.9 · $0.15GPT-6.1 SolClaude Opus 5.5GPT-5.6 SolClaude Fable 5.1Claude Fable 5Claude Opus 4.8
Scatter of Rams Index against cost per page for each model; filled points are on the frontier. The list beside it carries the same values.45505560657075$0.1$0.3$1$ per page (generation)GPT-6.1 Sol · 65.4 · $0.15Claude Opus 5.5 · 60.4 · $0.28GPT-5.6 Sol · 58.9 · $0.17Claude Fable 5.1 · 58.9 · $0.87Claude Fable 5 · 57.6 · $0.53Claude Opus 4.8 · 52.9 · $0.15GPT-6.1 SolOpus 5.5GPT-5.6 SolFable 5.1Fable 5Opus 4.8

A model is on the frontier when no cheaper model is clearly better. Clearly better uses the same paired interval as the rank, so a cheaper model has to beat it outright, not by a point inside the noise.

Cost is what the model's API charged to build the page, from the run ledger at list price on run day. Judging cost is ours and stays out.

  • GPT-6.1 Sol$0.15 · 65.4 · frontier
  • Claude Opus 4.8$0.15 · 52.9 · beaten by a cheaper model
  • GPT-5.6 Sol$0.17 · 58.9 · frontier
  • Claude Opus 5.5$0.28 · 60.4 · frontier
  • Claude Fable 5$0.53 · 57.6 · beaten by a cheaper model
  • Claude Fable 5.1$0.87 · 58.9 · beaten by a cheaper model

Head to head

Row model against column model.Each model against every model below it. The number is the paired difference in Index points.

Paired difference in Rams Index points, row model minus column model, with whether the 95% interval excludes zero.
GPT-6.1 SolOpus 5.5GPT-5.6 SolFable 5.1Fable 5Opus 4.8
GPT-6.1 Sol+5.0, inside the noise+6.4, inside the noise+6.4, clearly better+7.8, clearly better+12.5, clearly better
Claude Opus 5.5-5.0, inside the noise+1.4, inside the noise+1.4, inside the noise+2.8, inside the noise+7.5, clearly better
GPT-5.6 Sol-6.4, inside the noise-1.4, inside the noise0.0, inside the noise+1.3, inside the noise+6.1, clearly better
Claude Fable 5.1-6.4, clearly behind-1.4, inside the noise0.0, inside the noise+1.3, inside the noise+6.1, clearly better
Claude Fable 5-7.8, clearly behind-2.8, inside the noise-1.3, inside the noise-1.3, inside the noise+4.8, clearly better
Claude Opus 4.8-12.5, clearly behind-7.5, clearly behind-6.1, clearly behind-6.1, clearly behind-4.8, clearly behind

GPT-6.1 Sol

  • vs Claude Opus 5.5+5.0, inside the noise
  • vs GPT-5.6 Sol+6.4, inside the noise
  • vs Claude Fable 5.1+6.4, clearly better
  • vs Claude Fable 5+7.8, clearly better
  • vs Claude Opus 4.8+12.5, clearly better

Claude Opus 5.5

  • vs GPT-5.6 Sol+1.4, inside the noise
  • vs Claude Fable 5.1+1.4, inside the noise
  • vs Claude Fable 5+2.8, inside the noise
  • vs Claude Opus 4.8+7.5, clearly better

GPT-5.6 Sol

  • vs Claude Fable 5.10.0, inside the noise
  • vs Claude Fable 5+1.3, inside the noise
  • vs Claude Opus 4.8+6.1, clearly better

Claude Fable 5.1

  • vs Claude Fable 5+1.3, inside the noise
  • vs Claude Opus 4.8+6.1, clearly better

Claude Fable 5

  • vs Claude Opus 4.8+4.8, clearly better
  • Row clearly betterClearly better
  • Inside the noise
  • Row clearly behindClearly behind

Judges

The same pages through a second judge, GPT-5.6 Terra. The Index stays the Claude judge's score.

No sign the Claude judge favors Claude models. The Claude judge puts the Claude-built models 4.7 Index points behind the OpenAI-built ones; Terra puts them 4.5 Index points behind, on the same scale. The difference is -0.2 points (95% interval -3.7 to +3.6), so a lean of up to 3.7 points either way is not ruled out yet.

The two judges order the 6 models with a Kendall tau of 0.55 (95% interval -0.14 to 0.87). Across all 75 pages it is 0.54, and 0.31 within a brief, where the question is which build of the same page is better.

The widest split: Claude Opus 5.5 is second under the Claude judge and fifth under Terra.

Each model's order under the Claude judge, under Terra and under the two-judge panel, with the score behind it.
ModelClaude judgeTerraPanel
GPT-6.1 Sol165.4179.81+0.50
Claude Opus 5.5260.4570.74-0.05
GPT-5.6 Sol358.9279.32+0.16
Claude Fable 5.1358.9377.13+0.09
Claude Fable 5557.6474.55-0.06
Claude Opus 4.8652.9669.56-0.46
  • Order is by point score, 1 for the highest, not the interval rank above; most gaps between neighbors sit inside the noise. Each judge's score is its own Index, (grade sum − 7) / 28 × 100.
  • The judges use the scale differently (mean grade sum 28.0 for Terra, 23.4 for Claude), so each judge's Claude-minus-OpenAI gap is measured in its own spread across pages and shown in Claude Index points. The difference of the two gaps says whether the judges lean apart, not which one leans.
  • Panel is the mean of the two judges' grade sums, each standardized across the 75 pages, in standard deviations. Intervals: bootstrap, B=4000, resampling briefs, builds and passes. OpenAI Decisions has not judged model pages yet (v3 judges-2026-10-07 ran on websites and screenshots only).

Profile

The seven grades behind the Index, on the 1–5 scale (axis zoomed to the range used), mean of desktop and phone. Pick up to four models.

2.83.03.23.43.63.84.0HierarchyGPT-6.1 Sol · Hierarchy 3.80Claude Opus 5.5 · Hierarchy 3.60GPT-5.6 Sol · Hierarchy 3.40Claude Fable 5.1 · Hierarchy 3.45TypographyGPT-6.1 Sol · Typography 3.65Claude Opus 5.5 · Typography 3.40GPT-5.6 Sol · Typography 3.20Claude Fable 5.1 · Typography 3.35SpacingGPT-6.1 Sol · Spacing 3.25Claude Opus 5.5 · Spacing 3.15GPT-5.6 Sol · Spacing 3.18Claude Fable 5.1 · Spacing 3.10ColorGPT-6.1 Sol · Color 3.70Claude Opus 5.5 · Color 3.50GPT-5.6 Sol · Color 3.33Claude Fable 5.1 · Color 3.35ComponentsGPT-6.1 Sol · Components 3.60Claude Opus 5.5 · Components 3.40GPT-5.6 Sol · Components 3.57Claude Fable 5.1 · Components 3.35ComposureGPT-6.1 Sol · Composure 3.60Claude Opus 5.5 · Composure 3.45GPT-5.6 Sol · Composure 3.43Claude Fable 5.1 · Composure 3.50CraftGPT-6.1 Sol · Craft 3.70Claude Opus 5.5 · Craft 3.40GPT-5.6 Sol · Craft 3.38Claude Fable 5.1 · Craft 3.40
2.83.23.64.0HierarchyGPT-6.1 Sol · Hierarchy 3.80Claude Opus 5.5 · Hierarchy 3.60GPT-5.6 Sol · Hierarchy 3.40Claude Fable 5.1 · Hierarchy 3.45TypographyGPT-6.1 Sol · Typography 3.65Claude Opus 5.5 · Typography 3.40GPT-5.6 Sol · Typography 3.20Claude Fable 5.1 · Typography 3.35SpacingGPT-6.1 Sol · Spacing 3.25Claude Opus 5.5 · Spacing 3.15GPT-5.6 Sol · Spacing 3.18Claude Fable 5.1 · Spacing 3.10ColorGPT-6.1 Sol · Color 3.70Claude Opus 5.5 · Color 3.50GPT-5.6 Sol · Color 3.33Claude Fable 5.1 · Color 3.35ComponentsGPT-6.1 Sol · Components 3.60Claude Opus 5.5 · Components 3.40GPT-5.6 Sol · Components 3.57Claude Fable 5.1 · Components 3.35ComposureGPT-6.1 Sol · Composure 3.60Claude Opus 5.5 · Composure 3.45GPT-5.6 Sol · Composure 3.43Claude Fable 5.1 · Composure 3.50CraftGPT-6.1 Sol · Craft 3.70Claude Opus 5.5 · Craft 3.40GPT-5.6 Sol · Craft 3.38Claude Fable 5.1 · Craft 3.40

Rams Bench

The first track of the Index. Single-file pages built from written briefs, rendered and judged as a designer would see them.

Briefs
5

5 public briefs in v0. The private season-1 set (25 briefs plus these 5 as anchors) runs in v4.

Grades per pass
7 × 2

Hierarchy, typography, spacing, color, components, composure, craft. Each graded 1–5 at desktop and at phone size.

Pass-to-pass ICC
0.76

How closely two judge passes on the same page agree, on 15 pages judged twice. Visual score 0.75, Terra grades 0.91.

Bootstrap
B=4000

Resamples briefs, builds within a brief and passes within a page, paired across models.

  • Landing page built by GPT-6.1 Sol
    LandingGPT-6.1 Sol · 70.5
  • Dashboard page built by GPT-6.1 Sol
    DashboardGPT-6.1 Sol · 68.8
  • Settings page built by GPT-5.6 Sol
    SettingsGPT-5.6 Sol · 52.7
  • Pricing page built by GPT-6.1 Sol
    PricingGPT-6.1 Sol · 61.6
  • Article page built by GPT-6.1 Sol
    ArticleGPT-6.1 Sol · 78.6

Method

Everything the number depends on, stated once.

What is the Rams Index?

A leaderboard of how well models design interfaces. Each model builds the same pages from the same written briefs. Rams renders every page at 1440 and 390 px and grades it on seven dimensions a designer would look at. The Index is the sum of those grades, rescaled to 0–100.

Why the grade sum and not one overall score?

Seven separate grades repeat better than a single verdict. On real webpages the grade sum separates design-led, typical and rough sites more clearly than our finding-based visual score, and two passes on the same page agree more often. The visual score is still shown, as a secondary column.

How is rank decided?

Rank is 1 plus the number of models that are clearly better. Clearly better means the paired 95% interval on the difference sits above zero. Models whose difference sits inside the noise share a rank, so a tie is a finding, not an omission.

Does a Claude judge favor Claude models?

We run the same pages through GPT-5.6 Terra as a second judge and print it beside the score. No sign the Claude judge favors Claude models. The Claude judge puts the Claude-built models 4.7 Index points behind the OpenAI-built ones; Terra puts them 4.5 Index points behind, on the same scale. The difference is -0.2 points (95% interval -3.7 to +3.6), so a lean of up to 3.7 points either way is not ruled out yet. The Judges section shows each model's order under both. OpenAI Decisions (planned, v4) is the third judge from v4.

Who checks the judge?

Designers. This is season 0, published before their read: working designers rank the same pages blind, and the read rule was fixed before any of their answers were opened. Until that read (2026-11-06), this is the judge's opinion: stable and repeatable, not yet validated by designers.

What is private?

The brief texts of the season-1 set and the judge prompt. The method, the scores, the intervals and the sample pages are public. Five public briefs run alongside the private ones as anchors, so a model that overfits the public set shows up as a gap.

The full method, with every known weakness: How Rams measures design

Rams Index v0 (internal) · updated 2026-10-07 · engine f0bb659Run spend $20.03 · 180 calls · generation $12.92 · judging $7.12
Rams

Product

Company

Guides

Frameworks

Compare

© HSLA0001 Inc. 2026