Anthropic Launches Claude Fable 5, a State-of-the-Art AI Model with Safety Restrictions
Anthropic announced the release of Claude Fable 5, a highly capable AI model that achieves state-of-the-art performance on most benchmarks while including safeguards to prevent misuse in sensitive areas like cybersecurity. The company also launched Claude Mythos 5, an unrestricted version for authorized government and infrastructure partners through Project Glasswing. The models are priced at less than half the cost of previous versions, representing a significant advancement in AI capabilities paired with safety considerations.
Anthropic has released Claude Fable 5, described as a Mythos-class model with capabilities exceeding all previously available versions from the company. The model demonstrates exceptional performance in software engineering, knowledge work, vision, and scientific research, with particularly strong results on complex, lengthy tasks. To address safety concerns, Anthropic implemented safeguards that redirect sensitive queries to less capable models, with false positives occurring in less than 5% of sessions. Simultaneously, the company launched Claude Mythos 5, an identical underlying model with lifted safeguards for authorized users including cyberdefenders and infrastructure providers through Project Glasswing, a collaboration with the US Government. Early testing showed significant practical benefits, including Stripe reporting that the model compressed months of engineering work into days. Both models are priced at $10 per million input tokens and $50 per million output tokens, substantially less than previous versions.
What's missing
The announcement lacks independent verification of benchmark claims and does not address potential concerns about concentrating advanced AI capabilities with government partners or the long-term implications of tiered access models where some users receive unrestricted versions. Additionally, there is no discussion of how this pricing and capability structure might affect competition in the AI market.
How coverage differed
The source is Anthropic's official announcement, presenting the launch in promotional terms while emphasizing safety measures and real-world benefits. Independent coverage would likely scrutinize the effectiveness of safeguards, the implications of providing unrestricted versions to government entities, and competitive positioning relative to other AI developers.
What different sources said
- Hacker NewsCenter
Claude Fable 5
Related
Xbox's New CEO Prioritizes Gaming Over AI, Signals Return to Core Strengths
Asha Sharma, Xbox's new CEO since February, is refocusing the gaming division on its core gaming business rather than pursuing AI-driven initiatives, marking a strategic shift from her predecessor Phil Spencer. Sharma has implemented changes including lowering Game Pass prices, canceling AI features, and reviving exclusive franchises like Gears of War to reverse declining hardware sales and subscriber growth. Her approach signals Microsoft's recognition that Xbox needs to compete on gaming fundamentals rather than emerging technologies to regain market share against PlayStation and Nintendo.
Researchers Develop Ultrafast Machine Learning on FPGAs Using Kolmogorov-Arnold Networks
Researchers have designed hardware architectures for ultrafast machine learning inference and online learning using Kolmogorov-Arnold Networks (KAN) implemented on Field-Programmable Gate Arrays (FPGAs). FPGAs offer advantages over GPUs for applications requiring ultra-low latency and high hardware efficiency by implementing neural networks directly as digital logic rather than sequential processor instructions. This work addresses a gap in machine learning acceleration for specialized, latency-critical applications that cannot be efficiently served by traditional GPU-based approaches.
Nango's Evolution in Running Untrusted Customer Code: From Sandboxes to AWS Lambda
Nango, an API integration platform, has transitioned its approach to executing untrusted customer code from in-process sandboxes to distributed runners to AWS Lambda to improve security and resource isolation. The company processes over 150 million functions monthly across different workload types (on-demand calls, long-running jobs, and webhooks) while maintaining strict isolation requirements. This architectural evolution reflects the ongoing challenge of balancing security, cost, and performance when executing untrusted code at scale.