GPT‑6 Astra’s Pelican SVG Test Shows Superior Visual Output Over GPT‑5.6 Models

Experiment Setup

On the afternoon of 4 September 2026 the author gained access to the GPT‑6 Astra model and used it to create SVG images of pelicans riding bicycles.

Each image was generated at five reasoning settings—low, medium, high, xhigh and max—because Astra does not support a “reasoning=none” option.

The resulting SVGs were placed alongside comparable outputs from GPT‑5.6 Sol, Terra and Luna in a side‑by‑side comparison grid.

The grid was assembled using GPT‑5.6 Sol to render the visual layout, allowing a direct visual quality comparison across models and settings.

Findings on Quality and Cost

Visually, every Astra pelican, from low through xhigh, appeared clearer and more detailed than the best GPT‑5.6‑Sol result, which the author described as “a bunch of abstract shapes.”

The top‑performing GPT‑5.6‑Sol image (xhigh) still lacked the concrete form that even Astra’s low‑level output achieved.

Astra’s max‑level rendering was judged “really good,” though it occasionally missed rendering the pelican’s legs on both sides of the frame.

In monetary terms, Astra’s pricing is roughly double that of Sol, with input tokens costing $10 per million and output tokens $50 per million, compared with Sol’s $5 and $30 respectively.

Despite higher per‑token rates, Astra consumes significantly fewer tokens at each reasoning level, narrowing the effective cost gap between the two families.

Specifically, Astra and Luna each used 16 input tokens per request, while Sol and Terra required 26 input tokens.

This token efficiency means Astra’s low‑level generation, priced at about 9.55 cents, produces a superior pelican image than any Sol variant, which costs roughly 10 cents for a noticeably poorer result.

The author noted, “I wonder if Astra and Luna are more related to each other than OpenAI let on?” highlighting a possible architectural similarity.

For developers who need inexpensive yet high‑quality SVG graphics, the test suggests that even the cheapest Astra configuration may deliver better visual fidelity than higher‑priced Sol settings.

The observed token‑count disparity also implies that Astra’s underlying tokenizer may be more compact or that its prompt engineering requires fewer symbols to achieve the same effect.

Because Astra’s max‑level output still occasionally omits bilateral leg details, users requiring perfect symmetry might need to post‑process the SVG or select a higher reasoning level.

The broader implication is that newer generation models can offset higher headline pricing through reduced token consumption, delivering comparable or superior outcomes at similar effective costs.

Practitioners evaluating LLMs for image generation should therefore consider both per‑token pricing and actual token usage, rather than relying solely on advertised rates.

As model families evolve, side‑by‑side visual benchmarks like this pelican grid become valuable tools for making data‑driven selection decisions.

Why This Matters