Senior AI Journalist
OpenAI’s Astra Solves Ten Unsolved Math Problems, Signals New Era for Long-Running AI Tasks
OpenAI has made a significant leap forward in AI capabilities with its upcoming model family, Astra. This new system has reportedly solved ten previously unsolved mathematical problems across various complex fields, including high-dimensional geometry, coding theory, and quantum complexity. The breakthrough demonstrates Astra’s ability to tackle long-running and intricate tasks, moving beyond the short-term limitations of previous AI models.
The mathematical proofs generated by Astra were achieved with a surprisingly low computational cost, estimated at around $2,000 using existing API rates for OpenAI’s models. This efficiency, combined with the successful formalization of each proof in Lean—a system for creating machine-checkable certificates of mathematical correctness—underscores the potential for AI to accelerate scientific discovery significantly. OpenAI emphasized that while human researchers assisted in preparing the papers, the core mathematical arguments originated from Astra, highlighting a new paradigm for AI-assisted research.
Astra is designed to coordinate multiple AI agents over extended periods, addressing a critical weakness in current agentic systems: the tendency to compound errors during lengthy workflows. This development aligns with OpenAI’s long-term vision of creating fully autonomous AI researchers capable of working on complex problems for days or even centuries, far surpassing human capabilities. The system is also expected to be among the first to undergo testing under a new US regulatory framework for AI models, signaling a growing focus on safety and responsible deployment.
Source: The Decoder
Microsoft Shifts Away From OpenAI and Anthropic, Prioritizes In-House MAI Models for Cost Efficiency
Microsoft is making a strategic pivot in its AI strategy, gradually replacing models from OpenAI and Anthropic with its own internally developed MAI (Microsoft AI) models across several Copilot products, including Excel and Outlook. This move, reported by Bloomberg, aims to significantly reduce Microsoft’s reliance on third-party AI providers and cut down on associated costs, even as the in-house models currently handle a smaller fraction of total requests.
While Microsoft claims its MAI models are trained on “clean, commercially licensed data,” a technical paper revealed the use of the Common Crawl dataset, which has legally unsettled implications for AI training. This contrasts with Microsoft’s public stance of championing open-weight models and advocating against a few proprietary AI models dominating the industry. The transition raises questions about whether Copilot and Office customers might receive less capable AI for the same price, as Microsoft openly acknowledged its goal to “ultimately eliminate” the cost of external models.
The company’s head of AI, Mustafa Suleyman, indicated that AI billing could shift towards usage-based pricing, potentially making MAI models the default while offering third-party models as premium add-ons with surcharges. This approach would effectively pass on OpenAI and Anthropic expenses to customers. The ongoing shift underscores the intense competition and evolving cost structures within the AI ecosystem, as major tech players strive for greater control and efficiency in their AI offerings.
Source: The Decoder
Google DeepMind Undergoes Major Leadership Reshuffle as Hassabis Steps Back, Dean Departs
Google DeepMind is facing a significant leadership overhaul with CEO Demis Hassabis stepping back from daily operations to assume the role of Alphabet’s chief scientist. Simultaneously, Jeff Dean, a long-standing figure at Google, is leaving after 27 years to launch his own AI startup, Discovery Loop. These high-profile departures signal a new chapter for Google DeepMind as Koray Kavukcuoglu, the former CTO, takes the helm to lead the company in its race against top AI rivals.
Reports suggest that Hassabis’s decision to step back was driven by a preference for visionary scientific work over day-to-day management, a role he reportedly found unfulfilling. The departures also shed light on internal frustrations among Google researchers regarding limited access to the company’s powerful TPU chips. While Google Cloud sells these chips to competitors like Anthropic, its own research teams face constraints, highlighting a perceived bureaucratic hurdle that makes younger, more agile companies more attractive to top talent.
Despite these internal challenges, Google is actively expanding its AI empire. The company recently announced a substantial partnership with Mirendil, an “exciting frontier AI lab,” which will utilize over $100 million worth of TPUs and Nvidia GPUs through Google Cloud. This move demonstrates Google’s continued investment in the AI sector even amidst leadership changes and internal resource allocation concerns, indicating a dual strategy of nurturing internal talent while also fostering external partnerships.
Source: The Decoder
What to Read Next
- OpenAI’s Accidental Hugging Face Attack Timeline 2026: A Detailed Breakdown of the AI Security Breach
- Grammarly Go vs Sudowrite 2026: Advanced AI Writing Assistants for Creators & Authors
- AI Agents Go Rogue: Australia’s First Autonomous Cyber Attack Signals Urgent Safety Crisis
- How to Track AI Coding Costs at Scale in 2026: A Strategic Guide for Enterprise Teams
- Browse all AI Stack Digest articles
Bookmark aistackdigest.com for daily AI tools, reviews, and workflow guides.
This article was produced with the assistance of AI tools and reviewed by the AIStackDigest editorial team.