“It blows my mind”-“It has a tendency to overengineer things a little”: Developers react to road-testing OpenAI GPT‑5.6 Sol

OpenAI rolled out its GPT-5.6 family to app and API users worldwide at the start of July, with Sol positioned as the most capable model, Terra for general use, and Luna for speed. Sol was also pitched as the efficiency play, with OpenAI saying it could get “more useful work from every token” while running behind the company’s strongest safety stack so far.

That framing matters, but the early developer chatter is focusing less on the marketing and more on whether Sol actually helps finish work. Russell Twilligear, an AI research and development lead at BlogBuster in Dallas Fort Worth, says he has been comparing it directly with Anthropic’s Claude Opus 5. In his setup, Sol kept catching mistakes in a massive database project with more than three million entries, and he said the results were good enough to make him rethink where the lines are in this market.

There is a catch, though: Twilligear’s test was not a clean one-to-one evaluation. One model was creating while another was reviewing, which says as much about workflow as it does about raw model quality. Joshua Estrin, who runs AI-driven strategy and skills company Concepts in Success and studies model evaluation, makes that point bluntly. If one model misses errors in another model’s database, he says, that may reflect the prompt or the role more than a true head-to-head win.

The strongest praise comes when Sol is given room to think. New York-based Columbia Business School PhD candidate Shouqiao Wang said he solved six open Erdős problems in five days using GPT-5.6 Sol, with ultra reasoning effort and Codex doing the long-running work. He called it effective across a wider range of problems and better at sustained mathematical searches. But even he said Sol can overengineer things a little, and he still reaches for Claude or Gemini on some frontend tasks.

Edinburgh-based Viamki founder Jean Bustinza sounds similarly impressed, if not blinded by the hype. He calls Sol his go-to model, compares it to talking with a senior software developer, and says it is especially useful for broader architectural questions. Then the familiar complaint lands: sometimes it says something that makes no sense, and sometimes it overdoes the solution. That sounds less like a clean winner than a model that is very good at the hard stuff and occasionally gets a bit too clever for its own good.