GPT-6 Astra Tested: SVG Generation Quality Surpasses Entire GPT-5.6 Lineup

GPT-6 Astra's First Real-World Test

Well-known developer Simon Willison, upon gaining access to GPT-6 Astra, ran an interesting comparative experiment: asking different models to generate an SVG image of "a pelican riding a bicycle." He tested GPT-6 Astra across five reasoning levels — low, medium, high, xhigh, and max — and compared them against all three variants of GPT-5.6 (Sol, Terra, and Luna).

What seems like a lighthearted test actually reveals significant advances in the new model's graphic generation, reasoning efficiency, and cost-effectiveness. Through a side-by-side comparison grid, the advantages of GPT-6 Astra across multiple dimensions become immediately clear.

Generation Quality: A Clear Generational Leap

The results show that GPT-6 Astra achieves a qualitative leap in image generation:

Even at its lowest reasoning level, Astra's pelican output outperforms the best results GPT-5.6 Sol can produce at any reasoning level. Simon specifically noted that GPT-5.6 Sol's pelican — even at the xhigh level, which he considers better than max — still looks like "a collection of abstract shapes." By contrast, every output from Astra, from low to xhigh, is more realistic and coherent.

Astra at the max reasoning level is particularly impressive, generating a pelican with rich detail and a well-structured composition. There is one interesting caveat, however: below the max level, Astra still cannot reliably draw the pelican's legs on opposite sides of the bike frame — a detail that is only resolved at the highest reasoning level.

Cost-Effectiveness: Fewer Tokens, Better Value

In terms of pricing, GPT-6 Astra costs roughly twice as much as GPT-5.6 Sol: $10/million input tokens (vs. $5 for Sol) and $50/million output tokens (vs. $30 for Sol). But the real-world cost picture is far from a simple 2x multiplier.

The key finding lies in token efficiency: Astra consistently uses significantly fewer tokens than Sol at every reasoning level. This means that while the per-token price is higher, the actual cost gap for completing a given task is substantially narrowed.

The most compelling example is Astra at the low level: for just $0.0955, it generates a pelican image that beats anything GPT-5.6 Sol can produce at any level or price point. Spending a dime on any other model simply doesn't get you close. This cost-performance advantage is hard to ignore for applications that require large-scale image generation.

Technical Details: A Possible Connection Between Astra and Luna

Simon noticed an intriguing technical detail during his comparison: both Astra and Luna used 16 input tokens for the prompt, while Sol and Terra used 26.

This consistency raises an interesting hypothesis: could there be a deeper technical relationship between Astra and Luna than OpenAI has officially disclosed? While this is speculation based on limited data, the similarity in token usage patterns may hint at shared components in their tokenizers or underlying architectures.

Key Takeaways for Developers

This comparison offers several practical insights for AI application developers:

1. Choose the right reasoning level: For graphic generation tasks, Astra low may already meet the needs of most use cases — there's no need to default to the highest level.
2. Rethink cost budgets: Don't look at model unit price alone; token efficiency matters just as much.
3. Quality-first strategy: In scenarios where output quality is paramount, a newer model at a lower reasoning level may outperform an older model at its highest level.

This test of GPT-6 Astra highlights another important step forward in large language models' multimodal generation capabilities, and offers a concrete, intuitive reference for understanding performance differences across model generations.