In Peter Gostev's Arena comparison, GPT-6.2 Astra can produce a more convincing facade, while Claude Fable 5.2 can make the people around it feel like they belong there. GPT-6.2 Sol sometimes gets close enough for substantially less. The video moves from 3D worlds to HTML stories, storefronts, games and charts, with Claude Opus 5.5 joining selected comparisons. These are artifacts the models made with code, not just still images to judge at a glance.
Start with Sol for a straightforward prototype. Try Astra when geometry, lighting or fine visual detail is what's missing. Try Fable when you need a different creative direction, stronger scene pacing, or a fresh take on the interaction. Keep Opus in the comparison if it already works for you. The deciding metric is how much it costs to get a build you would actually use.
Watch Peter Gostev's Arena Comparison
The Astra Pricing Correction
The video's Astra estimates use older GPT-6 Astra pricing. Arena's note accompanying the video reports that the actual Astra 6.2 costs per generation were about 35% lower than the estimates shown using that older pricing. That is a correction to the video's run estimates, not a 35% cut in the API list rate. OpenAI lists standard input and output rates 20% lower than GPT-6 Astra, and cached input 60% lower. The mix of billed tokens matters; without run-level usage, the overlay is not a receipt.
| Model | Input / output per 1M tokens | What Gostev's demos suggest trying |
|---|---|---|
| GPT-6.2 Astra | $8 / $40 | Detailed geometry, lighting and precise visual execution. |
| GPT-6.2 Sol | $2 / $10 | Lower-cost prototypes and simpler explanations. |
| Claude Fable 5.2 | $8 / $40 | More varied art direction, scene movement and play mechanics. |
| Claude Opus 5.5 | $4 / $20 | A useful reference for visual identity; shown in selected tests only. |
Standard uncached API rates in USD as checked 18 November 2026, from OpenAI and Anthropic. Astra and Fable have the same base input and output rates; they can still produce different bills. Sol's rates are one-quarter of Astra's, but the ratio for any finished build can vary with reasoning effort, token usage, extra turns and rework.
3D Worlds: Surface Detail vs. a Scene That Feels Alive
Gostev prefers Fable's expansive Istanbul scene, but gives Astra the edge on facade detail in another example. In the parting-sea scene, he likes Astra's water and lighting while noting that Fable makes the people hurry through a frightening situation without being told to. The distinction shows up again in the HTML demos: Astra's domino sequence looks precise; Fable makes a car sequence livelier with camera changes. Neither treatment always wins. In the clay-style examples, Gostev likes Astra's materials and Fable's movement, and finds both capable in the crochet scenes.
For a real brief, name what you care about before comparing renders. A product walkthrough might need clarity and repeatable framing. A short explainer might need a camera change exactly when the viewer should notice something new. A single visual-quality score would hide that difference.
Websites and Games: Give Function Its Own Score
In the website round, Gostev sees Astra and Sol repeating invented branding, eyebrow text, subtitles and boxes across unrelated briefs. Opus varies its layouts more in some examples. Yet he also notes that the more distinctive Opus chocolatier design makes him scroll farther before reaching the purchase decision. Use the same test for both: can someone find and act on the primary task, and does the design feel like it belongs to this product?
The game tests make the split clearer. Gostev often prefers Astra's graphics and finds several control schemes solid. Fable sometimes feels snappier or gives the player a richer space to explore, but one Fable entry has a control bug that makes it unplayable. There is no clean winner in this set. Test movement, camera, restart and a complete gameplay loop before judging the screenshots.
Charts, Animated Explainers and the Peacock Outlier
Gostev likes the core charts in some Astra visualizations more than the generic headers around them. He also calls out inexpensive Sol runs for straightforward explanations and prefers Astra's treatment in parts of the historical and science animations. A clean chart or animated model can still be wrong: check units, labels, source data and the sequence of events before sharing it as an explanation.
The peacock is a different sort of result. Gostev kept asking Astra for refinements and reports a $171 on-screen estimate at the old rates. Fable got fewer turns, making this a demonstration of what Astra could reach with more iteration, not a fair one-shot comparison or a current bill. He likewise cautions against using the 100-dog exercise as a cost comparison because he tried different things, including motion.
Use the Prompts, Then Audit One Result
Browse Peter Gostev's 3D Prompt Collection or its JSON version. Those are Gostev's materials; it is a 3D collection, not a complete log of every live comparison. To test a smaller brief, copy the original JQ AI SYSTEMS prompt below. Use separate runs in the same setup; keep assets, tools and revision limits fixed and log each model's effort setting. Check first-frame quality, controls and desktop/mobile behavior before counting a run as accepted. For API tests, record elapsed time, token use and retry spend, then divide total spend by the number of builds you would use.
Build a small waterfront scene
One scene, fixed requirements, and clear checks.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Meet the models and the pricing context |
| 01:08 | 3D worlds: architecture, water and detail |
| 05:46 | Creative range vs. a consistent style |
| 11:48 | HTML stories and animations |
| 16:45 | Clay, crochet and emerging visual styles |
| 21:04 | Website UI: structure and repetitive layouts |
| 27:33 | Games: graphics, controls and bugs |
| 34:39 | Data visualization |
| 36:37 | Explaining concepts with animation |
| 38:41 | Pushing SVG detail: the peacock test |
| 40:10 | 100 ways to draw a dog |
| 41:29 | Final thoughts: Astra, Fable and Sol |
Sources and Limits
The Arena video and supplied transcript are the source for Gostev's examples and judgments. This is a curated hands-on comparison, not a blinded leaderboard or a review of production code. Not every model got the same number of attempts, and JQ AI SYSTEMS did not rerun the builds. You can try a brief yourself in Code Arena. Evidence cutoff: 18 November 2026.