Leaked sources say OpenAI aims to launch a new AI model called Astra next week, internally codenamed mewfour. According to the leak, Astra is the largest pre-trained model trained by OpenAI after GPT-4.5, marking a major leap in the company's model capabilities. Previously, it was reported that an internal version of Astra successfully solved 10 important open mathematical problems. If calculated based on Sol API rates, the total token cost required to find these solutions is approximately $2000, equivalent to about 13,500 RMB, demonstrating the model's impressive potential in complex mathematical reasoning.
Regarding its capabilities, according to a report from Bleeping Computer, Astra's core highlight lies in its ability to organize and coordinate multiple AI agents, making it suitable for long-term, high-difficulty tasks and oriented toward stronger reasoning and collaborative workflows. This design direction indicates that OpenAI is expanding its model capabilities from single-turn conversations and generation to multi-agent orchestration and complex task automation, aligning with the industry's current trend of transitioning from conversational AI to agent-based AI. If Astra indeed has agent collaboration capabilities, it could become a key infrastructure for enterprise-level AI applications.
The Accelerated Model Release Schedule: A Critical Move Under Competitive Pressure
OpenAI has recently significantly accelerated its model release schedule, and the timing of Astra's launch is closely related to the competitive pressure it faces. Google's flagship model Gemini 3.5 Pro has been delayed multiple times, while Anthropic's Claude series has been continuously eroding market share in programming and reasoning. OpenAI needs a new model to maintain its technological leadership. Releasing the signal for Astra's launch at this time is both a response to market expectations and an important measure to convey confidence to developers and enterprise customers.
Astra's breakthrough in solving mathematical problems is particularly noteworthy. Open mathematical problems have long been considered a hard benchmark for measuring AI reasoning capabilities. The ability to solve 10 such problems means the model has deep logical reasoning and multi-step inference capabilities, rather than simply providing approximate answers through pattern matching. Combined with its agent collaboration features, Astra may be positioned as a specialized model for scientific research, engineering, and complex decision-making scenarios, complementing the general conversation models of the GPT series. Whether OpenAI can widen the gap with its competitors will be determined by the official release next week.
