Nick Saraev opens his GPT-6.2 Astra review with disappointed reactions from other testers, then shows a mechanical watch page he genuinely likes. Both observations belong in the same review. Astra can produce a convincing visual prototype, but impressive first frames and low benchmark costs don't tell you what it takes to finish a project. Saraev's five builds put that difference on screen.
Try Astra for a tightly scoped visual build, especially where detail matters. Compare Sol when the brief is simpler and cost matters more. Keep your current model for open-ended work until Astra matches it on the quality of the finished result and how much help it needs to get there.
Watch Nick Saraev's Five-Build Review
What the Five Builds Actually Show
| Build | What Saraev showed | What to check before shipping |
|---|---|---|
| Watch page | A mechanical chronograph with an exploded view. Saraev preferred its detail to the GPT-6.1 Terra example he showed, a comparison with a smaller model. | Check that the parts explain the right product and the rest of the page supports the purchase decision. |
| 3D water page | A glass-of-water material study with lighting controls and a progressively refined render. The view needed time to sharpen after zooming. | Measure how quickly the scene becomes usable on a phone, not just how it looks after loading. |
| Arena shooter | A playable arcade prototype that Saraev enjoyed, with audio, enemy movement and a special ability. | Play a full round, including restart. Check control response and whether difficulty ramps coherently. |
| Robot escape | A 3D room with a working objective: restore power and reach the exit. The mouse cursor stayed visible, some visuals flashed, and on-screen directions gave away the short puzzle. | Test pointer capture and visual comfort. Let a fresh player try the puzzle without coaching. |
| Paris page | Architecture impressed Saraev more than the figures in the scene. | Inspect buildings at street level as well as from the opening camera angle. A plausible visual is not a surveyed reconstruction. |
In the robot run, Saraev describes Astra using GPT Image 2.6 for concept art before it builds the 3D scene. That's a result from a workflow that uses more than one tool. The video doesn't establish the origin or licensing of every asset, so verify both before using a demo commercially.
Cheaper per Benchmark Task Does Not Mean Cheaper per Finished Build
A Pareto frontier compares score and cost. A result lies on the frontier when no tested alternative scores at least as well for less money or scores higher at the same or lower cost. The frontier belongs to that particular benchmark and the configurations tested; change the task set or how you measure cost and it can move.
On the Artificial Analysis Intelligence Index, Astra at max effort scored 62 at about $4 per task. Fable 5.2 at max effort with default fallback scored 65 at about $26.54 per task. That is a real efficiency advantage on this suite, not a bill for any of Saraev's five builds. The same evaluator's launch report puts Astra first on its human-judged AA-3D evaluation, but finds it flat against GPT-6 Astra on the separate Coding Agent Index in Codex at max effort.
OpenAI lists standard API rates of $8 input / $40 output per million tokens for Astra and $2 / $10 for Sol. Cached input has separate rates. Saraev discusses Sol as the cheaper option but does not run it through the same five builds. Keep tool charges, retries and human cleanup in your comparison, or a cheap first pass can look better than it is.
What OpenAI's Simulation and CAD Demos Establish
The launch page shows a chassis-dynamics demo that combines a Python solver with 3D animation, and an RC buggy modeled from public photos and manuals. Those are OpenAI's examples, not Saraev's builds. They demonstrate that the model can connect several tools in a workflow. They don't validate the driving model against measured data or establish that the buggy design was manufactured. For a real project, check the technical assumptions and test the output separately from its presentation.
Run a Test That Matches the Work You Do
Give Astra and your incumbent the same small brief, allowed assets and revision budget. Before running it, decide what must work: say, a product that can be examined from three angles or a game that starts, reaches its objective and restarts without a page reload. Judge the finished artifacts before looking at their costs. Then record time, paid runs, review minutes and how many builds passed.
That's where Saraev's case for hands-on testing helps. A score can tell you how a model performed on a suite. It cannot tell you whether you enjoyed steering it, whether its visual choices fit your project, or whether someone else can use the result without you in the room.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | The reception so far |
| 00:57 | Build 1: watch landing page |
| 01:56 | Build 2: 3D water page |
| 03:04 | Build 3: arena shooter |
| 03:34 | Build 4: robot escape level |
| 04:59 | Build 5: Paris page |
| 05:35 | Benchmarks and cost per task |
| 08:28 | The OpenAI launch page |
| 11:05 | Why vibe benchmarks matter |
| 11:36 | Verdict |
Sources and Limits
The embedded video and supplied transcript establish Saraev's five demos and judgments. OpenAI is the source for its own launch examples and API rates; Artificial Analysis is the source for the separate evaluation scores and costs. The video does not report comparable bills for the individual builds. JQ AI SYSTEMS did not rerun them. Evidence cutoff: 18 November 2026.