This is a free preview of a subscribers-only post.
Was this newsletter forwarded to you? Sign up to get it in your inbox.
At the end of June, we lost access to GPT-5.6 Sol while it went through government review. Dan Shipper said returning to older models felt like going back to the Stone Age. Austin Tedesco compared GPT-5.5 to shooting a basketball twice as heavy.
Now Sol is back. It’s become our default for work we want to shape as we go: writing, coding, research, and analysis inside OpenAI’s new unified ChatGPT and Codex desktop app. Sol is fast, unusually good at finding and using context, and easy to redirect. Anthropic’s Fable still leads on broad assignments that require the model to decide what should be built in the first place.
We get into all of it in our Vibe Check, including:
- Why Sol scored 56/100 against Fable’s 90/100 on our Senior Engineer benchmark—and how 12,900 lines of unnecessary code explain most of the gap
- How Kieran Klaassen rebuilt Proof with Sol in about one-third the time Fable needed, only to prefer Fable’s design
- Why Sol finished last in our six-model writing benchmark while I (Katie Parrott) still used it to move through 24 drafts of this article in six to eight hours
- How Sol found Arielle Shipper’s email, inspected 46 CSV files, and returned with seven useful questions—then made a calculation error serious enough to shake her trust