Gemini 4 Carbon vs. Claude Fable 5.2: 103x Cheaper, but the Picture Is Worse

Choosing a flagship model by benchmark tables is already boring; all the numbers have long been side by side. So it's more interesting to watch what happens when two models get the same creative prompt and you tally the bill. The author of the Neurogen channel collected eight such comparisons of Gemini 4 Carbon and Claude Fable 5.2 from Twitter and edited them into a single video. This is not a lab study but a set of individual cases. Still, for a practitioner it is more useful than many official charts.

What Was Compared and What Came Out

The tasks vary: a voxel scene with a hero and rockets, a pixel-art city, a pirate ship, kinetic typography for a founders' pitch, an educational 3D animation about GPU warps, a Ghibli-style robot ruins scene, and a first-person lawn mower simulator.

The picture that emerges is fairly consistent:

• Fable holds a recognizable character and texture better. In the rocket test, Gru is recognizable by his goggles and scarf, while Carbon's scene collapsed into mush. On the robot ruins, Fable shows rust and moss, and in the lawn mower simulator the author describes the visuals as being "from another galaxy": the grass, the shadows, the mansion itself.
• Carbon is neater with diagrams and typography. The author found the explanatory animation about warps, with its flows and arrows, clearer than Fable's more voluminous but more decorative version.
• In some cases, such as the pirate ship, the difference comes down to taste.

Money and Time

The harshest number in the roundup comes from the lawn mower simulator. Fable spent $349.15 and 3 hours 32 minutes, while Carbon finished at $3.40 and 39 minutes. That is roughly 103 times cheaper and 5.3 times faster. In the pixel city, with the same ~60 minutes, Fable cost ~$18 and Carbon ~$6.2. In the rocket test it was $24.11 versus $53.34.

By per-token pricing, Carbon is four times cheaper: $2 per million input tokens and $10 for output, versus $8 and $40. But this is an introductory price valid until November 30. From December 1 it becomes $4 and $20, and the gap is cut in half. If you calculated your budget on the old rates, recalculate it.

"Cheaper per task" also calls for caution. If Carbon's result in the Gru case is mush, the $29 savings are worth nothing, because you're paying for an unusable result. The right metric here is the cost of an acceptable result, not the cost of a single run.

The time story is mixed. In two tests, Carbon took the same ~60 minutes as Fable at maximum reasoning, and the author notes it is slower than the previous Argon. Yet in the lawn mower case it was noticeably faster. Speed seems to depend heavily on the task and the reasoning settings.

How to Use This

The practical takeaway from the roundup's author sounds sensible:

• Need an expressive frame with a person, an object, or complex texture? Pay extra for Fable.
• Need an explainer, an infographic, an animated diagram, or many runs in a row? Carbon is cheaper and more predictable.
• If you have a pipeline with iterations (dozens of attempts until it works), a cheap model with a slightly lower ceiling often beats an expensive one.

A good strategy for teams: run prototypes and drafts on Carbon, and hand only the final render or frames where fidelity matters to Fable.

What to Keep in Mind

These are Twitter comparisons by different authors, not a controlled experiment. Each has their own prompts, reasoning settings, and criteria for "beautiful." Most cases involve a single run, with no statistics. In some tests prices aren't given at all, and the assessments of visual quality are subjective. The bonus case compares Carbon not with Fable but with the previous Opus, so it has little to do with the main debate. On top of that, Carbon's prices are temporary.

So don't change your stack based on one video. Take two or three of your own real tasks, run them on both models before the end of November, and count not just the cost but also the share of results you can show to a client. Those numbers will be your benchmark.