OpenAI Plans Year-End Model Releases Amid Safety and Technical Shifts
October 17, 2026 · 3 min read
OpenAI is expected to release at least one or two new artificial intelligence models by the end of the year. This development follows the company's decision to halt a weaker model for safety reasons and coincides with accidental disclosures regarding its underlying technology.
Employees indicate that the release timeline remains probable, though uncertainty persists regarding whether the upcoming launch will include the major flagship upgrade currently in development. The preparation for this new flagship model may require additional time. The system must undergo alignment to ensure its objectives match the expectations of researchers and customers. Furthermore, the software used to test the model's capabilities must be secured. Once leadership approves the safety evaluation results, OpenAI must provide the model to the United States government for testing up to thirty days before sharing it with trusted partners. This process is mandated by an executive order from President Trump regulating the evaluation of advanced artificial intelligence systems.
The next model utilizes Looped Transformers, a technique already employed in the current flagship model, Astra. Under the name Recurrent Depth, this architecture loops internal token representations through the same transformer layers multiple times before generating the next word. This approach significantly improves performance in mathematics and coding while reducing costs by allowing smaller models to perform like larger ones. However, this method shifts part of the reasoning process from readable text to internal numerical representations, making it invisible to human reviewers. This conflicts with chain-of-thought monitoring, where safety researchers read the model's output steps to control its behavior. Consequently, OpenAI has deliberately limited the use of loops in Astra to maintain a readable chain of thought.
Following an incident involving Hugging Face, OpenAI reportedly redirected approximately ten percent of its total computing resources toward alignment and safety work. This shift corresponds to hundreds of megawatts of data center capacity. The cost of renting this redirected capacity would amount to at least billions of dollars per year at current market prices. It remains unclear how many servers were previously or are currently dedicated to such work. Artificial intelligence researcher Noam Brown stated in a podcast in mid-September that ten percent of his team also works on security and safety.
Microsoft accidentally revealed details about Looped Transformer inference. The company, a major investor in OpenAI with access to its intellectual property, published a technical description of GPT-6.1 Sol. This model was launched earlier in the month as a cheaper alternative to Astra. The now-deleted section explained that GPT-6.1 Sol is based on looping inference, a form of reasoning where the model thinks through the same problem in multiple passes before generating an answer. The description stated that GPT-6.1 Sol uses the same base model weights as GPT-6 Sol but with two inference passes instead of three. A person with direct knowledge confirmed that the description accurately summarized information OpenAI had privately shared with Microsoft. Microsoft removed the passage after realizing it should have remained secret. The use of Looped Transformers is considered a sensitive aspect of OpenAI's model development, both in training and inference. OpenAI chief scientist Jakub Pachocki has acknowledged that monitoring chains of thought is increasingly fragile and moving in a negative direction, warning against an industry-wide race toward models without readable reasoning.