Can a model get cheaper and more capable without becoming a better daily driver? In Every's launch-day review, Dan Shipper says none of the team members who tried GPT-6.2 Astra switched to it. After roughly two weeks of early-access testing, the team saw a model that could research a coding task carefully and produce striking 3D detail, but that often needed more direction to finish work. The difference between can do it and will finish it without another round matters more than a headline score.
Keep your current model as the default until Astra 6.2 wins a task you actually repeat. Based on Every's tests, a clearly scoped change or a detailed 3D asset is a better first trial than moving all of your editorial work or open-ended app builds at once.
Watch Every's Two-Week Review
Disclosure: At 00:47, Shipper says Every receives early access and free test tokens from major labs, including OpenAI, and that the opinions are Every's own. The results describe the team's workflows, not a blind comparison across all professional tasks.
Writing: A Rewrite Isn't the Same as Editorial Feedback
Every asked both Astra versions to review the opening of Laura Entis's Why You Should Burn More Tokens. The brief put identifying the strongest idea and assessing the framing first, with replacement copy where useful. The older GPT-6 Astra led with its editorial diagnosis. GPT-6.2 Astra led with a new headline, subhead and opener, leaving the reasoning until later. Shipper found the newer opening so compressed that he had to reach the end of a sentence before understanding its point.
That is a narrower and more useful finding than declaring the model bad at all writing. A direct rewrite might suit a brief that asks for copy. An editor deciding whether to change an argument needs to see the diagnosis first. To retest it, ask for the strongest idea, one framing problem and the reason it matters. Ask for replacement copy last. Compare the feedback before comparing the prose.
Coding: Good Backlog Judgment, Uneven Follow-Through
At 05:47, Shipper reports that the two Astras scored about the same on Every's senior-engineer test, with Fable 5.2, Opus 5.5 and Sonnet 5.5 slightly ahead. On the junior-engineer task, 6.2 checked whether backlog work had already been done, selected a sensible item and researched it thoroughly. The team found Fable's resulting code easier to merge, though, and saw Astra stop early on several broader builds. That separates good preparation from a finished implementation.
Artificial Analysis found Astra flat against its predecessor on its own Coding Agent Index in Codex at max effort, even as the model improved on its broader Intelligence Index. Those are separate tests. OpenAI also reports that the new models omitted blockers more often in two evaluations and that Codex instructions reduced those omissions. That finding is about reporting blockers, not proof of why Every's builds stopped.
Design: Rotate the Camera Before Calling It Done
Tyler Nishida's lightly directed briefs exposed a trade-off: Astra had competent visual defaults, but Every saw less variety from one design to another than with Fable 5.2. In Every's Millennium Falcon test, 6.2 was the strongest OpenAI entry so far, according to Shipper, but not the overall winner.
Jack Cheng's miniature town makes the case for trying it on a specific job. Every describes Astra projecting generated 2D imagery onto 3D geometry to add visible detail. From one angle it looks richly textured; rotate the view and some of that apparent detail disappears. That's a useful rendering choice for a fixed shot. For an explorable scene, inspect the same object from several angles and make sure the geometry still meets the brief.
A Small Test Before You Change Models
OpenAI lists Astra at $8 input and $40 output per million tokens, 20% below the prior Astra's standard rates at launch. That makes a retest worthwhile, but it doesn't count review time. Run one task twice, once with your current model and once with Astra 6.2, using the same inputs and acceptance criteria. Judge the result before looking at the model name when you can.
- Writing: Can you name the argument and the suggested change without rereading?
- Coding: Did it finish the agreed change, run the relevant checks and list anything it couldn't verify?
- Design: Does the result work beyond the opening frame, and can the model produce a genuinely different alternative?
Track model spend, elapsed time and your own intervention time. Adopt it for the workflows where the finished result improves at a cost you can justify. A win on one workflow doesn't require a full migration.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Introduction to the two-week test |
| 00:29 | The quick take |
| 00:47 | Early access and token disclosure |
| 01:11 | Writing and editorial feedback |
| 05:47 | Coding benchmarks and end-to-end builds |
| 07:26 | Design defaults and 3D tests |
| 09:30 | What the release says about OpenAI |
| 11:00 | Read the full Vibe Check |
Sources and Limits
Every's video and the supplied transcript are the basis for the hands-on findings. Katie Parrott's written Vibe Check, published 17 November, covers both Astra and Sol; the full text is subscriber-only. This post focuses on the Astra video and does not present Sol as having gone through the same demos. JQ AI SYSTEMS did not rerun Every's tests. Evidence cutoff: 18 November 2026.