Tool Reviews

GPT 6.1 Sol: Astra-level intelligence at one-fifth the cost

Simon Willison's HN post ties GPT-6.1-Sol pelican examples to a near-Astra intelligence claim at one-fifth the cost.

GPT 6.1 Sol: Astra-level intelligence at one-fifth the cost

Key Takeaways

  • Simon Willison highlights GPT-6.1-Sol as offering “near-Astra intelligence for a fifth of the price,” based on OpenAI DevDay 2026 references and pelican-riding-a-bicycle examples.
  • The post is a quick visual signal rather than a formal model review: the pelican drawings show broadly similar output to the GPT-6 family, with no striking visual difference.
  • The pelican test is a lightweight sanity check for visual instruction following and coherent drawing output, not a measure of reasoning, coding, safety, or actual pricing.
  • The price-performance claim is the strongest takeaway, but the article does not provide benchmark scores, latency data, pricing tables, or task-specific evaluations.
  • Readers should wait for official DevDay 2026 materials, published pricing, and direct comparisons between GPT-6.1-Sol and Astra before making a model-selection decision.

Meta description: Simon Willison shared GPT-6.1-Sol pelican examples from OpenAI DevDay 2026, noting near-Astra intelligence at a fifth of the price on Hacker News.

GPT 6.1 Sol: Astra-level intelligence at one-fifth the cost

Tags: ai, openai, generative-ai, llms, pelican-riding-a-bicycle

Featured image: Abstract friendly illustration of a pelican drawing a bicycle beside a price tag; no real people, no logos, no text.

Simon Willison posted a Hacker News comment about GPT-6.1-Sol, using the phrase “near-Astra intelligence for a fifth of the price” and linking to pelican-riding-a-bicycle examples connected to OpenAI DevDay 2026. The post points back to a keynote liveblog dated Sept. 29, 2026, and Willison explains that the pelican examples were late because he had been live-blogging the event. For readers comparing large language models, the post is less a formal review than a quick visual signal: a model name, a price claim, and a familiar drawing test placed side by side.

The comment sits between keynote notes and visual checks

Willison’s post explains why the pelican examples arrived late: he had been live-blogging the OpenAI DevDay 2026 keynote. The source link points to a 2026/Sep/29 path for openai-devday-2026 on simonwillison.net. That detail matters because DevDay announcements often create a short window of noisy comparisons. A follow-up post that is explicitly tied to a liveblog gives readers a way to place the claim inside the event’s timeline rather than treating it as an isolated benchmark result.

GPT 6.1 Sol: Astra-level intelligence at one-fifth the cost

The Hacker News title names the model as GPT 6.1 Sol, while the body uses GPT-6.1-Sol. Keeping both forms visible is useful: model branding can shift between presentation slides, links, and code names. The comment then points to a Markdown SVG renderer at tools.simonwillison.net/markdown-svg-renderer, which is where the GPT-6.1-Sol pelicans are shown. A Markdown SVG renderer is a tool that converts Markdown text into rendered content, including vector graphics, so the pelicans here are likely a visual output check rather than a traditional image-generation gallery.

A second link points to a static GPT pelicans image on static.simonwillison.net, where Willison compares the newer examples with the GPT-6 family. That makes the post easy to scan: one link shows the new model’s output, another shows the broader family of pelican drawings used as a visual reference.

Pelicans are a lightweight model sanity test

The pelican-riding-a-bicycle prompt is a recurring visual benchmark in discussions of generative models and rendering pipelines. It asks a model or tool to produce a simple but structured drawing: a bird, a bicycle, and a plausible relationship between them. That makes it a practical sanity check. It is not a measure of reasoning, coding, safety, or price, but it can reveal whether a model or renderer can follow a visual instruction, keep shapes coherent, and produce something recognizable.

Willison notes that the GPT-6.1-Sol pelicans are “not notably different from the GPT-6 family pelicans.” The exact wording is important. The claim is not that GPT-6.1-Sol is identical to the GPT-6 family, nor that it failed the test. It says the visual difference is not striking. In model comparisons, “not notably different” can mean several things: the renderer is stable, the drawing prompt is too easy to separate models, or the new version has improved in ways that do not show up in this particular visual check. Readers should treat that as a limitation of the test, not proof that the model has no change.

The price claim gives the visual check more weight

The headline claim is stronger than the pelican observation: near-Astra intelligence for a fifth of the price. That phrase pairs two kinds of information that are often separated in AI coverage. Capability claims usually need benchmarks, task evaluations, and context windows. Price claims usually need published pricing, billing units, and real usage patterns. Willison’s comment brings them together in one line, which is useful for readers who need a quick comparison, but it is not a full evaluation report.

GPT 6.1 Sol: Astra-level intelligence at one-fifth the cost

For a developer, the practical question is whether a model that is described as near-Astra can handle the same work at a fifth of the price. That would change how teams choose between models for prototypes, batch jobs, chatbots, or content tools. A fifth of the price could make a model attractive for high-volume tasks, as long as quality, latency, safety, and reliability remain acceptable. The post does not provide benchmark numbers, latency data, or a pricing table, so readers should not infer those details from the pelican images or the one-line claim.

What readers should watch next

The most valuable part of this Hacker News comment is its placement: it links a keynote liveblog, a model name, a visual test, and a price phrase. For readers who follow OpenAI, the next step is to look for the official DevDay 2026 announcement materials, pricing details, and any published comparisons involving GPT-6.1-Sol and Astra. Those materials would clarify whether “near-Astra” refers to a named model, a benchmark category, or a shorthand used in Willison’s comment.

For readers who use models in products, the pelican gallery is a reminder to build simple checks that are easy to repeat. A drawing test can help spot rendering regressions, but it cannot replace task-specific evaluation. The more useful follow-up would be to compare GPT-6.1-Sol on the same prompts, logs, or workflows used for GPT-6, while also checking the actual cost per thousand tokens, per image, or per request once official pricing is available.

The post is short, but it gives a clear signal: GPT-6.1-Sol is being positioned as a lower-cost option whose visual output is broadly familiar, with a price-performance claim strong enough to attract attention on Hacker News. That combination is enough to put the model on a watchlist, not enough to make a final buying decision.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.