The Pelican comparison grid for Astra is pretty interesting

The author accessed GPT‑6 Astra and used it to generate SVG images of pelicans riding bicycles, testing five reasoning levels (low, medium, high, x‑high, and maximum). He then arranged those images in a comparative grid alongside outputs from the GPT‑5.6 models Sol, Terra, and Luna to assess differences in quality, token consumption, and cost.

Astra’s pelicans were markedly superior: even the low‑level version outperformed every Sol variant, while the maximum level achieved very high quality. At the lower levels, Astra still struggled to render the bird’s legs on either side of the composition. In terms of token usage, Astra and Luna required only 16 input tokens, compared with 26 for Sol and Terra, narrowing the price gap even though Astra’s cost per million tokens is roughly twice that of Sol (USD 10 for input and USD 50 for output versus USD 5 and USD 30). A low‑level Astra pelican cost just 9.55 cents, far less than the expense needed to obtain comparable results from the other models.

This experiment demonstrates how newer, larger models can deliver better visual quality while using tokens more efficiently, potentially reducing operational costs for image‑generation‑dependent applications. Understanding these distinctions is important for those evaluating generative‑AI alternatives in creative or commercial projects.

Read the original article on Simon Willison

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.