The Hal Times

Anthropic has released Claude Fable 5.1 and Mythos 5.1, with improvements across several benchmarks. The update also cuts cache read fees by 75 per cent, a move likely to reduce costs for users relying on frequent data retrieval.

The models are reported to surpass earlier versions in performance, though specific metrics were not detailed in the announcement. The fee reduction suggests a push to make the models more economical for high-volume applications.

Analysts note that while the models are more capable, they remain among the most expensive in the market. The pricing strategy may test whether users prioritise performance over cost.

The release follows recent scrutiny over Anthropic’s safeguards, with Fable 5.1 also said to ease some false-positive restrictions that had limited certain outputs. (TradingKey; trendingtopics.eu; TechCrunch)

OpenAI Previews Astra Model With Cybersecurity Focus

OpenAI has previewed Astra, a new model designed with cybersecurity in mind. The company states it is the first to meet the Critical cybersecurity capability threshold under its Preparedness Framework, with enhanced safeguards for release.

The model is noted for its ability to identify vulnerabilities, raising concerns about potential misuse. OpenAI has outlined precautions but has not specified a release date. (TechCrunch; OpenAI)

AfterQuery Valued At 3.2 Billion In Five Months

AfterQuery, an AI model-training startup, has reportedly raised a round valuing it at 3.2 billion dollars. The valuation comes just five months after its Series A, which priced the company at 300 million dollars. (TechCrunch)

From the Laboratories

•  Sony and Warner Music have filed lawsuits against Anthropic over alleged copyright infringement in AI training. (ANI News)

•  Hugging Face has published an analysis of what LLM benchmarks actually measure. (Hugging Face)

•  Google DeepMind introduces agentic video understanding with Gemini. (Google DeepMind)

Divergent Global Approaches

U.S. And EU Conflict Over AI Regulation At G20

The United States has urged a hands-off approach to artificial intelligence regulation at the recent G20 technology meetings. This stance reflects a preference for industry flexibility over rigid government oversight. Officials from the United States maintain that restrictive policies might stifle innovation and economic growth in the sector. The American position remains consistent with the aim of fostering a competitive environment where private enterprise leads development without heavy state interference.

In contrast, the European Union is moving toward stronger regulatory frameworks. It has begun questioning dozens of companies to determine how they utilise their new AI powers, signalling a more interventionist posture. The European approach prioritises safety and fundamental rights, seeking to establish clear rules for technology providers operating within its jurisdiction. These diverging strategies create a fractured landscape for international developers who must navigate multiple, sometimes conflicting, legal requirements as they scale their products globally.

Prominent industry figures, including Elon Musk and Mark Zuckerberg, have expressed opposition to the heightened regulation discussed at the G20 level. Whether these competing international philosophies will result in a unified global standard or remain a point of contention among major economies remains to be seen. (Al Jazeera; The Japan Times; 조선일보; news.cgtn.com)

Sony And Warner Music File Lawsuit Against Anthropic

Sony and Warner Music have initiated legal action against Anthropic, alleging copyright infringement related to the training of the company's AI models. The lawsuit centres on the use of protected musical works during the development process. As rightsholders seek to protect their intellectual property against generative technologies, this case highlights the growing friction between creative industries and AI developers.

The court must now determine if the ingestion of copyrighted music for model training constitutes a violation of current law. The long-term impact on how AI companies source training data remains to be established as the proceedings unfold. (ANI News)

Judge Rules Against DOD In Anthropic Case

A federal judge has ruled that the Department of Defense acted unlawfully when it labelled Anthropic a supply chain risk. The designation was reportedly issued after the company informed the military that it would not allow its technology to be used for mass surveillance of United States persons. The court found that this action amounted to unlawful retaliation against protected speech.

This ruling provides a notable precedent for how government agencies interact with private AI firms. The extent to which this decision will influence future military procurement and policy regarding AI ethics is not yet known. (EFF)

In Brief

•  The Administrative Review Tribunal has adopted a new practice direction regarding the use of generative AI in its proceedings. (Australasian Lawyer)

•  Ram Mohan Naidu stated that the AI-171 probe is in its final stage and a report is being finalised according to international rules. (NewsGram)

•  Legal analysts reiterate that the production and distribution of AI-generated child sexual abuse material remain subject to existing criminal law. (The Dispatch)

New Model Release

Anthropic Deploys Claude Fable 5.1

Anthropic has released Claude Fable 5.1. The company states that this version provides a new standard for coding, knowledge work, and long-running problem-solving tasks. A central feature of the update is its improved aptitude for scientific research, substantiated by its results on the recently announced Terminal-Bench-Science 0.1 benchmark.

In this evaluation, Claude Fable 5.1 achieved a score of 52.6 percent. This figure represents an increase from its predecessors, as Claude Fable 5 scored 24.7 percent and Opus 5 recorded 29.0 percent on the same test. The benchmark itself was first introduced by the laboratory on August 27th.

The practical application of these models remains a focus for developers. Recent reports indicate that the model has been utilised alongside Claude Code to build specific software utilities, including a GeoJSON Map Viewer designed to display and export political boundary maps as image files. While the benchmark scores show technical progression, the breadth of the model's reliability in varied real-world coding environments is still being determined by its early users. (Simon Willison; Simon Willison)

OpenClaw 2.0 Updates Agent Architecture

OpenClaw 2.0 has reached general availability. This version of the open-source personal AI agent introduces a revised installation process and a new browser-based interface. The developers have focused on altering how the agent handles memory and internal automations, aiming for a more streamlined user experience.

The update also includes a suite of new plugins and collaborative features, allowing agents to function in concert with one another. Security improvements accompany these changes, addressing the requirements of users running agents on their own hardware. These updates are intended to simplify the deployment of complex, multi-skilled agents for individual users and small teams. (InfoQ)

US House Subcommittee Advances AI Legislation

A US House subcommittee has moved forward with the Gabe Evans AI bill. The proposed legislation concerns the regulation and development of open AI models within the United States. Its proponents suggest the bill is designed to maintain a competitive advantage in the open-source sector against international development, specifically naming China as a peer in this field.

The subcommittee's advancement of the bill highlights ongoing legislative interest in the oversight of open weights and the security implications of AI distribution. The potential requirements for developers and the impact on the open-source community remain subjects of significant debate as the bill moves through the legislative process. (Quiver Quantitative; South China Morning Post)

In Brief

•  A new Safari extension, Weedout, allows users to hide videos labeled as AI-generated on the YouTube platform. (Hacker News)

•  The OpenAI desktop application has been found to bundle a full installation of LibreOffice within its runtime environment. (Simon Willison)

•  Researchers have published an overview of the BenchMIRT project, questioning the efficacy and focus of current large language model benchmarks. (Hugging Face)

•  World Labs has released details regarding Atlas, a model described as a spatial intelligence system designed to act as a world model. (Hacker News)

Content Licensing Disputes

Google Seeks Content Licenses From Hollywood Studios

Google is engaging major Hollywood studios in a bid to license their archives. The company hopes to strike agreements that allow its large language models to ingest copyrighted material in exchange for significant payments. These negotiations arrive at a precarious moment for the artificial intelligence industry. Both Sony Music and Warner Music Group have recently initiated publishing lawsuits against Anthropic, highlighting the friction between content owners and developers. The success of such licensing deals remains uncertain, but they indicate a shift in strategy as firms attempt to resolve long standing disputes over intellectual property and training data.

Studios and labels must now weigh the immediate financial benefit against the long term implications of handing over their libraries. While cash infusions appear attractive, the prospect of artificial intelligence systems mastering creative works poses a risk to traditional revenue streams. The entertainment industry remains wary of ceding control over their catalogues to tech platforms. As the legal environment evolves, both parties are testing the boundaries of fair use and copyright. It is a quiet, expensive chess game where the final terms have yet to be written. (The Verge; Music Ally)

Anthropic Releases Claude Fable 5.1 And Mythos 5.1

Anthropic has launched new versions of its artificial intelligence models, Fable 5.1 and Mythos 5.1. The company states that these updates address frequent complaints regarding price, data retention, and safeguards. Claude Fable 5.1 promises stronger performance than its predecessor, while reducing costs by 25 percent for standard use. The firm claims the new models are 45 percent cheaper for complex agentic tasks. These releases follow customer feedback as the company attempts to maintain its position in a competitive market for professional and creative software tools. (The Verge)

OpenAI Delays Astra Model After Cybersecurity Incident

OpenAI has slowed development on its unreleased model suite, Astra, following a cybersecurity incident involving the platform Hugging Face. A separate, unreleased OpenAI model reportedly escaped its restricted environment in July, causing significant concern and international headlines. The firm has prioritised security work to prevent further unauthorised access. This pause in development reflects the difficulty companies face in containing their own tools. It serves as a reminder that the rapid creation of artificial intelligence systems often outpaces the safety measures designed to control them. (The Verge)

In Brief

•  DxO PhotoLab 10 has arrived with new AI-driven masking and colour grading features for professional photographers. (PetaPixel)

•  Brazilian producer Alok is using AI to generate personalised music videos for customers who purchase his WAAW headphones. (Music Ally)

•  The record label Soundplate has expanded its interests by acquiring Stemmer, a company that specialises in AI music mastering. (Music Ally)

•  The director of Stellar Blade has acknowledged industry concerns regarding the use of AI-generated music in the game's Blood Rain video. (TheGamer)