GPT-6 Astra

OpenAI, closed weights, released Sep 2026, run at provider defaults through OpenRouter, served by OpenAI. 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

66.4

95% interval 57.7–75.0

Rank: 1of 14 models by Index; the paired intervals allow 1 to 6
Cost per page: $0.43Generation only, list price on run day
Second judge: 80.7GPT-5.6 Terra places it second of 14. The Claude judge: first.
Grade sum: 25.60Seven 1–5 grades, out of 35
Pages judged: 1010 judge passes, 2 builds per brief
Defects per page: 6.5Text under 12px (10 of 10)

Rank 1 of 14; the data allows anywhere from 1 to 6. Within noise of Grok 4.7, GPT-6.1 Sol, Claude Sonnet 5.5, Gemini 4 Argon and Muse Wonder: the gap to each sits inside the noise, so the board can't order them yet. Not clearly behind any model. Clearly ahead of GPT-6 Luna, Claude Opus 5.5, GPT-5.6 Sol, Claude Fable 5.1, Claude Fable 5, Gemini 3.8 Flash, DeepSeek V4 Pro and Claude Opus 4.8.

Designer check pending · read Nov 6Season 0.5New this seasonProvisional: 1.0 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.

Hierarchy3.851 of 14
Typography3.503 of 14
Spacing3.301 of 14
Color3.653 of 14
Components3.801 of 14
Composure3.801 of 14
Craft3.702 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.

Landing72.31 of 14
Dashboard62.53 of 14
Settings50.98 of 14
Pricing65.22 of 14
Article81.21 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: 113sMedian over every build, broken ones included
Output tokens: 7.7kMedian per page, reasoning included
Reasoning tokens: 0.5kMedian per page, as the API reports them
Broken pages: 0 of 10Every build came back as a whole page
Criticals: 90%9 of 10 scored pages had a critical issue in their first Rams code review
Built through: OpenRouterServed by OpenAI

Best and worst page

Of its 10 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: Article82.1
Index on this page, one judge pass. Brief mean 81.2.
- Single type register throughout bodyModerate
- Sidebar TOC vertical rhythm breaks below itemsModerate
- Eyebrow and field note label share baseline but differ in case/weight treatmentModerate
Worst: Settings50.0
Index on this page, one judge pass. Brief mean 50.9.
- Navigation tabs clipped at right edgeSerious
- Key count badge competes with headingModerate
- Footer tagline and user card vertically far from navModerate

All 10 of its pages sit beside every other model's in the gallery.

Head to head

GPT-6 Astra against every other model: the paired difference in Index points, and whether the 95% interval on it clears zero.

vs Grok 4.7+0.4Inside the noise
vs GPT-6.1 Sol+1.1Inside the noise
vs Claude Sonnet 5.5+4.6Inside the noise
vs GPT-6 Luna+4.6Clearly better
vs Gemini 4 Argon+4.8Inside the noise
vs Muse Wonder+5.9Inside the noise
vs Claude Opus 5.5+6.1Clearly better
vs GPT-5.6 Sol+7.5Clearly better
vs Claude Fable 5.1+7.5Clearly better
vs Claude Fable 5+8.8Clearly better
vs Gemini 3.8 Flash+10.9Clearly better
vs DeepSeek V4 Pro+11.8Clearly better
vs Claude Opus 4.8+13.6Clearly 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