GPT-5.6 Sol

OpenAI, closed weights, released Jun 2026, run at provider defaults through the OpenAI API. It built the same five pages as every model on season 0.5 of the Rams Index, and Rams graded each one at desktop and phone size.

Rams Index, 0 to 100

58.9

95% interval 51.6–66.9

Rank: 9of 14 models by Index; the paired intervals allow 3 to 11
Cost per page: $0.17Generation only, list price on run day
Second judge: 79.3GPT-5.6 Terra places it fourth of 14. The Claude judge: ninth.
Grade sum: 23.50Seven 1–5 grades, out of 35
Pages judged: 1520 judge passes, 3 builds per brief
Defects per page: 4.7Text under 12px (15 of 15)

Rank 9 of 14; the data allows anywhere from 3 to 11. Within noise of GPT-6.1 Sol, Claude Sonnet 5.5, GPT-6 Luna, Gemini 4 Argon, Muse Wonder, Claude Opus 5.5, Claude Fable 5.1 and Claude Fable 5: the gap to each sits inside the noise, so the board can't order them yet. Clearly behind GPT-6 Astra and Grok 4.7. Clearly ahead of Gemini 3.8 Flash, DeepSeek V4 Pro and Claude Opus 4.8.

Designer check pending · read Nov 6Season 0.5Provisional: 1.3 judge passes per page, protocol 2

Grades by dimension

The seven grades behind the Index, 1 to 5, mean of desktop and phone. The tick is the mean across all 14 models.

Hierarchy3.4010 of 14
Typography3.2012 of 14
Spacing3.186 of 14
Color3.3311 of 14
Components3.574 of 14
Composure3.439 of 14
Craft3.389 of 14

Bars start at 2.8 and end at 4.2, the range the board uses, so small differences show.

Index by brief

The same Index on each of the five briefs, and where the model places on that brief among all 14.

Landing67.06 of 14
Dashboard52.111 of 14
Settings52.74 of 14
Pricing51.27 of 14
Article71.710 of 14

Bars run from 40 to 90.

How it ran

Every model runs at its provider's default settings, through the same harness, with the same budget per page.

Effort: DefaultThe provider’s default. We send no effort or thinking setting
Time per page: 57sMedian over every build, broken ones included
Output tokens: 8.5kMedian per page, reasoning included
Reasoning tokens: 0.2kMedian per page, as the API reports them
Broken pages: 0 of 15Every build came back as a whole page
Criticals: 33%5 of 15 scored pages had a critical issue in their first Rams code review
Built through: OpenAI APIServed by OpenAI

Best and worst page

Of its 15 judged pages, the one the judge graded highest and the one it graded lowest, with the judge's top findings on each, most severe first.

Best: Article75.0
Index on this page, one judge pass. Brief mean 71.7.
- Meta line breaks inconsistently across two rowsSerious
- Byline meta uses en-dash rendered as bullet separator inconsistencyModerate
- Blockquote quotation mark orphaned above ruleModerate
Worst: Dashboard46.4
Index on this page, one judge pass. Brief mean 52.1.
- Duplicate 'Analytics' label: page title and active nav stateSerious
- Four identically-treated stat cards: no differentiation, no hierarchy among metricsSerious
- Duplicate total figure in chart headerModerate

All 15 of its pages sit beside every other model's in the gallery.

Head to head

GPT-5.6 Sol against every other model: the paired difference in Index points, and whether the 95% interval on it clears zero.

vs GPT-6 Astra-7.5Clearly behind
vs Grok 4.7-7.1Clearly behind
vs GPT-6.1 Sol-6.4Inside the noise
vs Claude Sonnet 5.5-2.9Inside the noise
vs GPT-6 Luna-2.9Inside the noise
vs Gemini 4 Argon-2.7Inside the noise
vs Muse Wonder-1.6Inside the noise
vs Claude Opus 5.5-1.4Inside the noise
vs Claude Fable 5.10.0Inside the noise
vs Claude Fable 5+1.3Inside the noise
vs Gemini 3.8 Flash+3.4Clearly better
vs DeepSeek V4 Pro+4.3Clearly better
vs Claude Opus 4.8+6.1Clearly better

The other models

Season 0.5 of the Rams Index, in board order, then the unranked.

Rams Index v0.5 · updated 2026-10-17Judged by Rams eyes · Claude Sonnet 4.6