Edgar Cervantes / Android Authority
Add Android Authority on Google:
TL;DR
- OpenAI is introducing GPT-6.1 Sol, promising Astra-like performance at one-fifth the expenditure.
- Across numerous benchmarks, GPT-6.1 Sol comes close to 6 Astra’s capabilities, and outpaces GPT-6 Sol.
- GPT-6.1 demonstrates fewer factual errors than GPT-6 Sol across similar tests.
DevDay is here for OpenAI, and that means the company is sharing a whole bunch of news about its latest work on models and the services built around them. While we just heard that GPT-6.1 Astra is getting sent back to the bench for misbehaving, progress continues on the rest of OpenAI’s headlining models, and this afternoon we’re getting the scoop on the first member of that 6.1-series to actually make its public debut, with the launch of GPT‑6.1 Sol.
While we only just got GPT-6 Sol, OpenAI is not a company to rest on its laurels, and today is pushing right on forward with the evolution of GPT-6.1 Sol. Even though Astra remains the more capable model, with these 6.1 improvements, OpenAI reports that Sol is rapidly approaching Astra in “intelligence,” nearly matching the flagship model in performance across multiple metrics.
In coding, GPT-6.1 Sol is neck-and-neck with 6 Astra, and also performs competitively at answering professional questions and working with documents — all at around one-fifth the cost of Astra. This general trend continues across evaluations like computer usage and scientific research, with GPT-6.1 Sol either directly outperforming previous models, or offering comparable performance at increased speed or with decreased expense.
Accuracy is the elephant in the room with far too many AI models, and even some of the best will still confidently report incorrect information. Compared to GPT-6 Sol, 6.1 pushes lower-effort response rates containing factual errors down from 11.4% to 7.7%. And once again, 6.1 Sol manages to perform comparably to 6 Astra here, albeit it for pennies on the dollar.
Unlike the obstreperous GPT-6.1 Astra model that was caught deceiving researchers, Sol didn’t raise any red flags during safety evaluations, so hopefully we don’t have much to worry about there.
Access is getting started across OpenAI plans with ChatGPT Work and Codex, but not Chat — at least, not yet. OpenAI also teases that it plans to share GPT‑6.1 Sol Ultrafast in the days to come, with nearly an order of magnitude improvement in token generation speed.
Thank you for being part of our community. Read our Comment Policy before posting.
