Less than 24 hours after announcing that it won’t release its new model, GPT-6.1 Astra, because it’s too dangerous to make public, OpenAI showed up to its annual DevDay event with a fresh release after all—just an apparently less powerful one. On Tuesday, the company released GPT-6.1 Sol, an update that promises Astra-level performance at a fraction of the price.
If you feel like you’re starting to lose track of OpenAI’s models, you’re forgiven. This summer, when OpenAI first launched GPT-5.6, it split it into three different models: Sol, Terra, and Luna. Practically speaking, its really a branding exercise. Sol is the powerful one, Terra is the mid-tier option, and Luna is the lightweight version.
When the company released GPT-6 earlier this month, it called it Astra and promised that it was its most capable model for intelligence work. Then it released two other versions of GPT-6, Luna and Sol. Terra is nowhere to be found and Sol is no longer the high-powered model. And now, with the release of GPT-6.1 Sol, it’s supposed to be comparable to Astra in how well it does with coding and intelligence work tasks, but have a significantly lower token cost. Maybe OpenAI could ask ChatGPT how to establish a meaningful brand identity.
Anyway, 6.1 Sol surpasses the GPT-6 version of the model in complex tasks like programming, debugging, and executing multi-step business workflows, while also reportedly reducing its error rate from previous versions. It does this while costing just $0.10 per million tokens, according to OpenAI, which it claims is bout half of what GPT-6 Sol costed for the same type of inputs.
The company claims Sol approaches GPT-6 Astra performance in some of those areas, but Astra remains the company’s most powerful model to date. In fact, if you take OpenAI’s word for it, GPT-6.1 Astra is so powerful that it wouldn’t be safe to make it available to the public. The company initially planned to release that model during DevDay, but announced on Monday that it was shelving it because the team was worried the model had reportedly “regressed” in safety and would do things like perform tasks without human permission.
GPT-6.1 Sol apparently doesn’t have these problems, so it got the green light for DevDay instead. Per OpenAI, GPT-6.1 will be available starting Tuesday for all users with a Plus, Pro, Business, Enterprise, or Edu subscription to ChatGPT Work and Codex. The model isn’t available yet in Chat, but one would assume that’ll be in the pipeline for the future.
Read the full article here
