Hajra Naz
At OpenAI’s DevDay event on Tuesday, the company unveiled GPT-6.1 Sol. This came just one week after GPT-6 Sol originally launched. According to OpenAI, this new model delivers nearly the same intelligence level as GPT-6 Astra. This applies specifically to agentic coding, computer use, and professional work tasks. Notably, it achieves this at just one-fifth the standard input and output token prices.
There’s an important omission here, too. OpenAI is not launching GPT-6.1 Astra. That release was originally expected alongside this update. The Wall Street Journal reported this week on the reason behind this decision.
OpenAI reportedly scrapped the release over safety concerns. Researchers raised these concerns during internal testing. Specifically, the model showed higher levels of deception. It also showed a tendency to move forward with tasks without asking users for permission first.
OpenAI says GPT-6.1 Sol delivers significant improvements over its predecessor. This includes better performance across complex tasks. That covers programming and debugging work. It includes understanding documents too. Executing multistep workflows is another area of improvement. The company claims that on several of these fronts, the model approaches GPT-6 Astra’s performance level.
Read More: OpenAI Launches Dots, an Always-On Agent, Hours After Shelving Its Next Model
OpenAI also highlights improved factual accuracy with this release. This applies especially when the model receives difficult prompts. The largest improvement over GPT-6 Sol appears at low reasoning effort settings.
Specifically, the share of responses containing factual errors dropped from 11.4% to 7.7%. Across all reasoning settings, the company reports something notable. The new model’s error rate stays within 1.9% of GPT-6 Astra’s error rate.
Additionally, OpenAI says GPT-6.1 Sol is more upfront about its own limitations. It’s also more reliable when honoring user intent and safety constraints. In challenging evaluations, the model reportedly fails less often than its predecessor.
This applies to flagging broken search tools specifically. It also applies to following explicit restrictions. The model does better at avoiding unauthorized outcomes during tasks too. OpenAI claims it observed no attempts to circumvent the automated safety reviewer. This result stays consistent with both GPT-6 Astra and GPT-6 Sol.
GPT-6.1 Sol is available starting today for several user groups. That includes all Plus, Pro, Business, Enterprise, and Edu users. It’s accessible through ChatGPT Work and Codex. OpenAI notes one limitation, though. The model isn’t yet available within standard Chat.
