Computer use
Give it a goal and it opens the page, fills the form, runs the flow, and hands the result back.
64px tall on desktop (4rem; 54px below 1020px), fixed over the canvas. Content sits in a 1440px container with 32px gutters: wordmark 17px, nav links 13px / 500 with 16px side padding, the 40px search button closing the nav group, then the glass "Log in" pill and the white "Try" pill, both 36px tall with 20px side padding and 12px icons. Transparent at the top of the page; once scrolled, a solid #000000 surface fades in over 400ms cubic-bezier(.33,1,.68,1). No blur.
GPT
Astra
Scroll
Split labels at 64px / 500, letter reveal 1s cubic-bezier(.22,1,.36,1) with 0.85–1.25s stagger, ambient glow #23435f. The particle canvas sits under this layer.
We’re introducing a model that brings together years of research across pre-training, reasoning, and alignment.
It reads long documents, understands screenshots, writes code that runs, and asks the right follow-up question. Inline code looks like astra.setScroll(); shortcuts look like ⌘ K.
Read the system card for methodology, or see all benchmarks
Give it a goal and it opens the page, fills the form, runs the flow, and hands the result back.
Reads the whole repository, the whole contract, the whole thread, and keeps the thread.
Refuses the same way every time and explains why in one sentence.
Solid #1f1f1f, glass #ffffff1f, or hairline outline. No shadows: depth comes from fill and glow.
Long documents, screenshots, code that runs, a follow-up question asked at the right moment.
These are not features bolted together; they are the same stars standing in different shapes at different times.
676px column, h2 48px / 500, p1 17px / 28px in #ffffff99. Nothing on either side: the star rails live there.
It takes action
Give it a goal and it opens the page, fills the form, runs the flow, and hands the result back to you.
576px × 80vh frame (shortened here). The SVG is invisible in production; it only reserves the aspect ratio for the stars.
| Terminal-Bench 4.0 | Score |
|---|---|
| This model | 91.4% |
| Previous generation | 72.1% |
| Comparison A | 65.0% |
| Comparison B | 48.7% |
h5 labels, tabular figures, white for the featured row, white-at-alpha bars. No colour.
Hairline pill container with 4px inner padding; items are 40px pills at 14px / 500, the selected one filled #ffffff1f. Autoplay: every 6s the next tab is selected; inside the selected pill a #ffffff33 fill holds for 2s, then grows left-to-right (scaleX, linear) over 4s. Hover, leaving the viewport or hiding the page pauses it; a click switches immediately and restarts. Reduced motion: no fill, no autoplay. The chart card below is the linked panel.
Sits in the 669px prose column: 16px radius, 1px #ffffff33 border, black at 20%, padding 16px 20px 20px. Title 18px / 500, a 32px select pill (#333333 at 80%, 1px #ffffff33 ring, 14px / 400, 12px side padding) that opens a dropdown (#2b2b2b menu, 12px radius, 8px padding, 40px options, selected #ffffff1f), a 20px download icon with no hover fill that opens a right-aligned action menu (#2b2b2b, 12px padding, 32px rows: Download PNG / SVG, Copy link, all functional), a legend with shape markers, a dashed reference line, mono tick labels, and an italic centred caption. Series colours come from the chart palette (blue-3 #539af8, blue-2 #2c67c5, orange-2 #ac4f23, orange-1 #653218), never from UI tokens.
“Astra gives us a significant advantage in both capability and efficiency. It successfully executes our most complex creative workflows while using up to 20% fewer tokens than other models we’ve tested.”
899px solid #1f1f1f card with a #ffffff1f border, 6px radius, 32px vertical and 24px horizontal padding. Logo row 32px, quote at h5 size / 500 with -0.025em tracking, attribution 14px / 500 in #ffffff99. Slides stack and crossfade over 250ms cubic-bezier(.23,1,.32,1); the nav is two 32px round buttons and a tabular “n / total” counter; a 1.5px ring around “next” fills linearly over the 5s auto-advance (fades in over 1250ms, out over 250ms) and pauses on hover, focus or a hidden page. Quotes are fixture text.
Worked for 13m 15s ›
Your personal website is live and private.
Working for 20s
I’ll use your background to build a site that supports the move. One question first: what role are you moving into?
Media (screenshots, video, side-by-side product UI) sits in a 10px-radius frame with a soft hairline on #0a0a0a; nothing inside gets a shadow. Column labels at h4 size, centred.
| Alignment | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Gemini 3.8 Flash |
|---|---|---|---|---|
| Internal computer use safety benchmark (lower is better) | 2.4% | 22.0% | 9.5% | - |
| Internal circumvention benchmark (lower is better) | 0.00% | 0.29% | - | - |
| ExploitGym honeypot (lower is better) | 0.0% | 48.2% | - | - |
| Internal hallucination benchmark (lower is better) | 4.2% | 12.2% | - | - |
Fixed layout, first column 34%, 20px row padding, hairline #ffffff33 above every row and below the last, 17px / 400 cells, 500 header, missing values as a muted dash. No zebra stripes, no highlighted column.
On ARC-AGI-3 the featured model was run with a responses harness that keeps past reasoning. ↩
Third-party scores are the providers’ reported numbers where a run was not possible. ↩
Chip: 32px radius, #ffffff1f, 8px 12px, 14px / 500. Footnotes: 14px list on a two-column inline grid (12px number, 8px gap, 12px between entries), heading 14px / 400 in #ffffff99, back-links unstyled. Logo strip: 44px buttons, 8px padding, 14px in #ffffff99, 4px radius, the active one white. Media bar: 14px / 500 tabular timestamps (white | 40% white), 32px square 4px-radius buttons, 2px track.
669px wide with a 959/540 aspect ratio, #1f1f1f, 4px radius. The stage keeps 24px side and 18px vertical padding around a white 16:9 slide that crossfades over 250ms. The 56px bar has a 1px #ffffff33 top rule, 12px side padding and a 1fr-auto-1fr grid: thumbnails toggle left, previous / “n / total” (56px, 14px tabular) / next in the centre, fullscreen right; buttons are 32px round with 16px icons, hover #ffffff99, disabled #5d5d5d. The thumbnails sidebar (132px, #404040, 100×56 white thumbs with a 1px #ffffff33 border and a 20px numbered badge, the current one ringed by a 2px white outline offset 2px) slides open over 200ms cubic-bezier(.33,1,.68,1). Arrow keys move; fullscreen uses the Fullscreen API.
1. Measured on a held-out set; see the system card for methodology.
2. Availability varies by region.