This afternoon I gained access to the GPT‑6 Astra model and used it to generate SVG images of pelicans riding bicycles. I ran generations at multiple reasoning levels — low, medium, high, xhigh and max (Astra does not support reasoning=none) — and arranged the results in a comparison grid alongside GPT‑5.6 models: Sol, Terra and Luna. The grid, which contains the full‑quality images, was useful for visually highlighting differences across models and settings.
Notable differences
- Overall quality: GPT‑6 Astra’s pelicans were clearly higher quality than those produced by the GPT‑5.6 family. Even Astra’s lower reasoning levels produced images that surpassed the best GPT‑5.6 Sol attempts. The Astra max output was noted as particularly strong.
- Consistency: Astra below max still sometimes failed to reliably place pelican legs on both sides of the frame.
- Model variation: the best Sol pelican (the author preferred xhigh to max) remained fairly abstract in form, while every Astra pelican from low to xhigh looked better than that Sol example.
Cost and token usage
The published pricing and token observations are:
- Astra: roughly $10 per million input tokens and $50 per million output tokens.
- Sol (GPT‑5.6): $5 per million input and $30 per million output. Although Astra appears more expensive per token, it used significantly fewer tokens at each tested level, which narrows the effective price gap between models.
A concrete data point: Astra low produced a better pelican than any GPT‑5.6 Sol model at any level for a cost of 9.55 cents. Spending 10 cents on any other model yielded a noticeably worse image.
Input token counts were also notable: Astra and Luna each used 16 input tokens, while Sol and Terra used 26. That raises the question whether Astra and Luna might be more closely related than publicly described by OpenAI.
Additional notes
- The author shared the transcript used to create the GPT‑6 Nova pelicans, and the full comparison grid with the high‑quality images is available in the original post.
- The test was primarily visual/creative but proved informative: it was fun and also revealed measurable differences in image fidelity and token efficiency across models.
Conclusion
This hands‑on comparison indicates that GPT‑6 Astra is a meaningful improvement over GPT‑5.6 models for SVG generation of figurative scenes such as pelicans on bicycles. Astra delivers better quality even at lower reasoning settings and tends to use fewer input tokens, though it is pricier per token. The token‑use similarity with Luna suggests further investigation into model relationships would be worthwhile.
