GPT-6.2 Soul vs Astral vs Gemini 4: Aesthetics Still Trail Fable

Three-task test: GPT-6.2 Soul ≈ Astral in quality, but Soul is faster and cheaper — the value winner.
Based on a Bilibili creator's three-task test, this article compares GPT-6.2 Soul, Astral, Gemini 4 Carbon, and Argon on speed, price, and frontend aesthetics, using official harnesses with medium reasoning effort. Soul and Astral outputs look nearly identical, but Soul is faster and cheaper; Argon ≈ Carbon at lower cost. Fable 5.2 still leads on aesthetics but costs more and runs slower. The final verdict hinges on Soul's Plus quota.
Three-Task Test: GPT-6.2 Soul vs Astral — Speed, Price & Aesthetics Compared
GPT recently released two new models: 6.2 Soul and Astral, while the Gemini camp brought out Gemini 4 Carbon and Argon. A Bilibili creator ran a three-task hands-on comparison covering speed, price, and frontend output quality. This article is based on that video, cross-referenced with earlier Gemini and Fable series tests, to answer one question: Is the GPT-6.2 upgrade worth switching for?
Test Setup: Same Official Harness, Medium Reasoning Effort
The creator used both vendors' official harnesses, all with medium reasoning effort, Skills disabled, and identical MB files to keep the comparison fair. Models tested:
- GPT-6.2 Soul (referred to as A or B — the video speculates repeatedly)
- GPT Astral (A or B)
- Gemini 4 Carbon
- Gemini 4 Argon
The video also revisits Gemini 4 Hydrogen, DeepSeekAd V4.1 Flash, and Fable 5.2, which had been tested earlier with the same prompts, providing a cross-model reference. Note that the creator never fully confirmed which GPT model corresponds to A and B, so some ranking conclusions are partly speculative.
Harness here refers to the official evaluation or invocation framework provided by model vendors, typically including unified prompt templates, tool-calling interfaces, and a result rendering environment. Using the same harness eliminates "frontend scaffolding differences" as much as possible, letting different models run through identical input and output pipelines — making results more comparable than running each model through its own third-party UI.
Reasoning effort is a parameter introduced by reasoning models that controls the token budget or number of steps for internal reasoning before generating the final answer. Medium effort balances quality and latency, suitable for frontend generation tasks that need some planning but not deep reasoning. Setting it uniformly to medium prevents any model from gaining an unfair quality advantage due to a higher default.
Task 1 — "Jelly Island": Soul Beats Astral on Both Speed and Price
The first task, "Jelly Island," required the models to generate an interactive scene with a pool and houses.
Speed and Style Differences
Gemini 4 Carbon took about an hour, while both GPT models finished in around half an hour. Carbon's output style differed noticeably from the previous Gemini 4 generation — the creator described it as more "blocky," though the water texture was actually quite good; only the animation speed was too fast, giving a visually "accelerated" feel.
Both GPT models leaned toward the traditional "GPT style" in color, with gradients that were no longer soft transitions but crisp color switches. The house models were oversized (a common longstanding issue), but overall presentable. The creator guessed in the video that contestant B might be Soul — a judgment later partly confirmed.

Aesthetic Ranking and the Price Reversal
The creator's aesthetic ranking was: A first, C second, B third. A was confirmed to be GPT-6.2 Soul, C was Gemini 4 Carbon. Although Soul's aesthetics were less refined than A (Gemini 4 Carbon), it held a crushing advantage in speed and price — Soul is cheaper than Astral and faster too.
The creator exclaimed repeatedly: "This Soul is impressive, folks — it's even faster than Astral." From a practical standpoint, if Soul's quality is close to Astral's, it's clearly the better deal.
The A/B/C labels that recur throughout are tags the creator assigned to each model's output during blind testing — used to rank aesthetics without knowing model identities, then mapped back to specific models after the reveal (e.g., A = GPT-6.2 Soul, B = Gemini 4 Carbon). Blind evaluation reduces brand bias, but it has a limitation: when a label's model identity isn't fully confirmed, ranking conclusions can only be treated as speculation. Readers should distinguish between the "blind ranking" and the "post-reveal value conclusion" — the latter is the core judgment of this article.
Task 2 — "Character Riding a Bicycle": Gemini 4 Argon and Carbon Hard to Separate
The second task was animation: generating a character riding a bicycle.
Animation Fluidity and Aesthetic Divergence
Gemini 4 Carbon's animation was fairly fluid, but the creator considered its "aesthetics average." Fable 5.2's version showed a distinct watercolor style — the character waved, expressions changed, and it pulled ahead aesthetically. GPT Astral's version was mocked: "its arms are worse than an elementary schooler's drawing."
Gemini 4 Argon took about six minutes, Carbon over ten — yet the creator found the results "really about the same." The video's conclusion was blunt: Argon is good enough.

Ranking and Cost Considerations
In this round the creator ranked: A first, C second, B third (because B's legs were detached from its body). But the more important takeaway is on price — Soul is cheaper than Astral, and Argon is far cheaper than Carbon within the Gemini lineup. For everyday use, the creator recommends prioritizing Soul and Argon.
Task 3 — "Island": Carbon Wins on Texture, Soul and Astral Indistinguishable
The third task was the island scene shown earlier, focusing on water reflections and overall texture.
Output Comparison
Gemini 4 Carbon immediately showed strong reflective texture; panning around, the water reflections held up well. The creator judged B (Gemini 4 Carbon) clearly better than A, while A and C were stylistically similar and hard to rank. This round's ranking: B first, C second.

Soul vs Astral: Virtually Zero Visible Difference
The key conclusion came after the prices were revealed: Soul is notably cheaper than Astral and faster too. The creator said outright: "Tell me — what's the difference between this Soul and Astral? Besides being cheaper, it's just as fast, right? Can any of you tell them apart?" In the outputs, the sky cloud gradients and light bulb details were essentially indistinguishable. So this round: Soul wins outright.
API cost and subscription quotas are two different billing models. API charges per token or per call — suited to developers and batch tasks, with costs scaling linearly with usage; subscription plans like Plus typically cap messages or credits, throttling or downgrading once exceeded. When the article says Fable 5.2 has "high API cost," it means higher unit price or consumption under pay-as-you-go billing; the closing concern about "how much Soul quota is available to Plus users" is a subscription allocation issue. The two aren't directly convertible, which is why the "which model is cheaper" conclusion depends on context: heavy developers look at API unit price, regular users look at usable quota within their plan.
Cross-Model Reference: DeepSeek Holds Its Own vs GPT-6.2, Fable 5.2 Crowned on Aesthetics
The second half of the video revisited other models tested with the same prompts, providing a broader frame of reference.
DeepSeek V4.1 Flash Performs Impressively
On the island task, DeepSeek V4.1 Flash's lighting was on par with the GPT-6.2 group. Its animation physics were reasonable and colors decent. The creator felt that "honestly, GPT-6.2 wasn't that much better than DeepSeek this round." DeepSeek was, however, slower.
Gemini 4 Hydrogen: Lightweight and Fast
Hydrogen was described as "lazy" but extremely fast — producing results in twelve minutes. This fits the small-model positioning: lightweight, fast, adequate quality.

Fable 5.2: Aesthetic Champion, But High API Cost
Fable 5.2's island was called the "aesthetic champion" by the creator — the lighting and texture were genuinely beautiful. Across all comparisons, Fable 5.2 still leads in frontend aesthetics. The trade-off: noticeably higher API cost and slower speed (the island task took twenty-six minutes).
Conclusion: Soul Is the Value Pick — But Plus Quota Is the Decider
Across the three tasks, the creator's overall take: GPT-6.2 Soul and Astral are barely different; Soul brought the speed up and is still the cheapest. Likewise, Gemini 4 Argon and Carbon differ little — Argon suffices for daily use.
The real gap is in aesthetics: Fable 5.2 remains the strongest but costs more; for value, GPT-6.2 Soul and Gemini 4 Argon are the more practical choices.
The video ends on an open question: if Soul's quota for Plus users is generous, this upgrade is decent value; if the Plus allocation is small, the conclusion may be the same as last time — just go use Gemini 4 Flash.
For developers, the selection logic boils down to: choose Fable for aesthetics first, Soul or Argon for speed and cost first, and for complex modeling and animation wait for more systematic future tests. The creator also plans follow-ups on modeling and more general prompts — worth watching.
One-line Summary
Three-task test shows GPT-6.2 Soul and Astral are nearly indistinguishable in quality, but Soul is faster and cheaper — the value winner.
Paragraph Summary
Based on a Bilibili creator's three-task test video, this article compares GPT-6.2 Soul, Astral, Gemini 4 Carbon, and Argon on speed, price, and frontend aesthetics. The test used official harnesses, medium reasoning effort, and Skills disabled for fairness. The three tasks were a jelly island, a bicycle animation, and an island scene. Results: Soul and Astral outputs are virtually identical to the eye, but Soul is faster and cheaper; Gemini 4 Argon and Carbon perform similarly with Argon more economical. On aesthetics, Fable 5.2 remains strongest but is costlier and slower, while DeepSeek V4.1 Flash holds its own vs GPT-6.2 but is slower. The final verdict hinges on Soul's quota allocation for Plus users.
Related articles

Amazon's First Digital Photo Frame Is 50% Off for Black Friday: Is the Alexa Frame Worth Buying?
Amazon's first digital photo frame, the Alexa Photo Frame, is nearly 50% off for Black Friday, with prices starting at $114. We analyze why this new product dropped in price just one month after launch, its positioning, who it's for, and buying advice.

Bob's Burgers' Linda Voice Actor John Roberts: Finishing One Thing Is Happiness
Bob's Burgers' Linda voice actor John Roberts discusses the Season 16 finale and shares his "happy to finish one thing" professional mindset, analyzing the voice stability behind long-running animation and a comic outlook on life.

The Road to Orbit: How Launch Megastructures Will End the Rocket Era
From mass drivers and skyhooks to launch loops and orbital rings, SFIA's long-form video maps the launch megastructure spectrum and shows how traffic volume drives technology choice.