Amir Mushich reports GPT-6.2 Astra cost $3.12 in a Blender comparison
Amir Mushich reports spending about $3.12 with GPT-6.2 Astra versus $19 with Claude Fable 5.2 on the same Blender modeling and animation task. He preferred Astra's first model, but found neither result satisfactory.

TL;DR
- A Blender modeling and animation task cost about $3.12 with GPT-6.2 Astra versus $19 with Claude Fable 5.2, according to AmirMushich's comparison.
- Astra's first version was his favorite, but none of the outputs satisfied him as a designer, he said in his verdict.
- A second Astra pass brought its two-session total to 31.3 million tokens and improved the render while worsening the animation, he reported in the same thread.
Astra sent back a draft before building out the whole system, Mushich reported. The hand's wires were Bézier curves with round bevels, a Blender technique that turns a curve into a 3D tube.
One hand, two xHigh runs
Designer Amir Mushich ran the hand test in his own Blender installation on an RTX 4070 laptop. He set both models to xHigh effort and asked them to recreate and animate the same reference hand at similar render quality.
Tokens and API costs
Mushich reported these figures for Astra's first run and the Fable run:
| Reported usage | GPT-6.2 Astra | Claude Fable 5.2 |
| --- | ---: | ---: |
| Total tokens | 3.5 million | 21.2 million |
| Output tokens | 26,000 | 236,000 |
| Approximate API cost | $3.12 | $19 |
Fable spent roughly six times as much. Its output went into an elaborate procedural build, including circuit boards and actuators, plus additional checks, according to Mushich.
He said 98% of Fable's tokens were cached context reread at each step. Anthropic reuses a stored prompt prefix on a cache hit, as its prompt-caching documentation describes.
Fable 5.2's listed API rates put cache hits at $0.20 per million tokens and output at $40 per million. Those token totals therefore combine work billed at very different rates.
Modeling, motion and rendering
Mushich judged the outputs on three separate dimensions:
- Modeling: Fable's hand looked “way worse.”
- Movement logic: Fable had a slight edge.
- Rendering: He slightly preferred Fable's finish.
He nevertheless picked Astra v1 as the best of the three outputs, while rejecting all three in his final assessment. A slight rendering win couldn't rescue the hand.
Astra v2
A second Astra session brought its cumulative usage to 31.3 million tokens, exceeding Fable's reported total. That changes the first-run efficiency comparison: the two-session Astra figure covers an additional iteration, and Mushich did not give a revised dollar cost.
He said the revision improved render quality but made the animation “complete trash.”
Local rendering time
Astra v1 rendered locally in about 21 minutes; Fable took 1.5 hours. Mushich excluded rendering time from his comparison because the laptop was otherwise free during Astra's render and busy during Fable's.