Anthropic has upgraded Claude Fable 5.2 and Claude Mythos 5.2, models designed to tackle the most daunting technical challenges. These models show significant improvements in areas like coding, reasoning, and vision. During prerelease testing, they assisted engineers in completing otherwise-infeasible software migrations, designing three working electronic devices from scratch, solving two open problems in black hole physics, and formalizing the proof of Fermat's Last Theorem in just 30 hours.

Claude Fable 5.2 is available to all users with strong cybersecurity and biology safeguards to prevent misuse. Claude Mythos 5.2, on the other hand, is accessible to organizations pursuing beneficial applications in cybersecurity or biomedical research through Anthropic's expanded Life Sciences and Cyber Verification Programs. These new models generate outputs up to 30% faster than their predecessors and are available at a lower price, with an optional fast mode that speeds up token generation by up to 2.5 times.

Extensive prerelease safety testing, including novel evaluations for concerning behaviors like sandbox escapes, was conducted to ensure safety and alignment. The models performed as well as, or better than, recent models on most alignment measures, with high-stakes risks from misalignment assessed as low. Benchmarks demonstrate that Fable 5.2 and Mythos 5.2 surpass Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon across nearly all evaluated capabilities, including agentic coding, scientific research, cybersecurity, protein design, and robotics.

Their coding prowess is remarkable, with over 90% of changes merged into Anthropic's core codebase now written by Claude. Fable 5.2 also designed a working operating system by coordinating over 1,000 subagents and migrated a critical production service from Python to Rust. In cybersecurity, Claude Mythos 5.2 is Anthropic's most capable model, catching more vulnerabilities and being better at reproducing and fixing them than previous models.

In science and engineering, Fable 5.2 solved two open problems in theoretical physics, while Mythos 5.2 found a novel way to accurately predict whole capsid structures for gene therapies by building on open-source protein-folding models. In mathematics, a team of 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in just 30 hours with minimal human intervention.

For hardware and robotics, Fable 5.2 designed and produced three working electronic devices and showed substantial advances in controlling physical robots, even for tasks it was not explicitly trained on. In research and analysis, early testers found that Fable 5.2 can handle more ambiguous, data-heavy tasks and exhibits better judgment, particularly in abstract reasoning and design.

Strong protections against illicit distillation have been implemented, reducing data harvested by distillers by 90% to 99%. Additionally, Enterprise Frontier Safeguards enable automated safety monitoring on prompts and responses stored within a customer's infrastructure, compatible with zero data retention (ZDR). Claude Fable 5.2 is now available across all major platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure.

👉 Read the full article on Anthropic

Google has expanded its eighth-generation TPU portfolio with the introduction of TPU 8tx for secure, confidential inference and TPU 8i Max for massive context windows. These new systems are designed to manage complex agentic workloads by providing specialized hardware for security, density, and reasoning. They join the existing TPU 8t for large-scale training and TPU 8i for real-time reasoning, offering specialized performance for every stage of the AI lifecycle.

The TPU 8tx integrates hardware-level security directly, protecting and isolating sensitive data and model weights even during active use. This makes it ideal for regulated enterprises and governments running frontier or proprietary models and autonomous agents with their most sensitive data. The TPU 8i Max, conversely, is engineered to address the memory wall bottleneck by combining two low-latency TPU 8i chips into a single package, delivering double the compute, SRAM, and HBM bandwidth of TPU 8i.

Google is enhancing TPU accessibility by supporting PyTorch, open-source software, and automated tools for developers. This includes open-sourcing TorchTPU, providing out-of-the-box TPU backends for leading open-source inference engines like vLLM and SGLang, and introducing a new open-source TPU agent suite that integrates with coding tools like Claude Code and Cursor to automate TPU onboarding and optimization.

The Agent Compute Engine software automatically suspends and resumes idle sessions for stateful AI agents, significantly cutting operational infrastructure costs by eliminating idle compute waste. Furthermore, TPU Sync, a model-state layer for TPU clusters, unifies memory, storage, and networking into a single, co-designed memory hierarchy, breaking through the memory wall by automatically tiering massive KV caches off-chip in milliseconds.

Now generally available on Google Kubernetes Engine (GKE), Cloud Pathways brings the Google DeepMind runtime behind Gemini to every Google Cloud customer, enabling orchestration of large-scale TPU systems from a single controller. Additionally, GKE dynamic slicing, utilizing optical circuit switching (OCS) technology, decouples physical hardware provisioning from logical slice formation, reducing job startup times by up to 5x and accelerating fault recovery by up to 4.5x.

Google's co-design approach, involving silicon architects, networking engineers, compiler teams, and Google DeepMind researchers, leverages AI tools like AlphaChip and Gemini to rapidly accelerate the development of next-generation TPU hardware.

👉 Read the full article on Google Blog

The AI hardware startup Nominal has seen its valuation triple to an impressive $3 billion in just eight months. This rapid increase in valuation highlights strong investor interest and high expectations for the company's technology and market potential within the AI hardware sector.

This news reflects the current trend where the advancement of AI technology is driving significant investment in hardware infrastructure, suggesting the growing importance of specialized hardware, especially with the increasing demand for AI agents and large language models.

👉 Read the full article on TechCrunch