Anthropic Unveils New Tier Above Opus: Claude Fable 5
On June 9, 2026, Anthropic released a new tier, Claude Fable 5 (model ID: claude-fable-5), upending its traditional three-tier lineup of Opus / Sonnet / Haiku. Arriving just 12 days after the release of Opus 4.8, it is positioned as a Mythos-class model.
Japan's tech community reacted immediately: Zenn already hosts articles rounding up its pricing, API usage, and Amazon Bedrock support. Comparison tests against GPT-5.5 are also underway, with some reports noting that on RDB implementation tasks it can still miss fine details even at the medium setting. Practical insights are being shared as well, including cost comparisons for adopting Claude Code in real-world workflows and a roundup of 10 Claude Code configurations for architects.
Apple Releases AI Inference Framework Core AI
Apple has released a new AI inference framework, CoreAI. Its distinguishing feature is that you can convert any AI/LLM model of your choice into the CoreAI format and run it on iOS and macOS, and Apple also provides an official CoreAI-based inference engine. On Qiita, an article on building an OSS implementation right after the release has been drawing attention.
Core AI could make local AI inference on Apple platforms far more accessible, and AI adoption on edge devices is expected to accelerate further.
Meta Partners with Reliance to Build Its First AI Data Center in India
Meta has signed a deal with India's Reliance for a 168-megawatt AI data center. This marks Meta's first AI data center investment in India. The facility is designed to support Meta's global AI computing demand and to allow for future expansion.
With its demographic dividend and advancing digitalization, India's AI demand is expanding rapidly, and this investment underscores the strategic importance of the market.
Kuaishou Publishes Technical Report on Vision-Language Model Keye-VL-2.0
Kuaishou (快手) has published a technical report on its vision-language model Keye-VL-2.0 on Hugging Face Daily Papers. Research announcements from Chinese companies keep coming in the multimodal AI space, and further progress is expected in integrating visual understanding and language generation.
Interpreting and Controlling Text-to-Speech with Sparse Autoencoders
Research published on Hugging Face Daily Papers demonstrates that applying sparse autoencoders (SAEs) to text-to-speech models makes it possible to precisely edit generated output. TTS models such as CosyVoice3 have so far been limited to control via prompts and predefined tags, but steering SAE features is said to enable much finer-grained control over speech generation.
The study also confirmed the existence of features that activate on audio alone, text alone, and both — a finding drawing attention as a step toward a deeper understanding of the models' internal representations.
Other News
Generative AI Headwinds for Management Consulting**: According to ITmedia, bankruptcies and closures among management consulting firms are hitting a record pace amid the rise of generative AI. Analysts note that the hardest hit are firms that fail to differentiate through expertise and depend on labor-intensive business models
Design Philosophy in AI Development**: Zenn carried multiple articles rethinking how AI development itself should be approached, including the argument that "the more you put AI at the center, the more you drift off course" and a report on a Scott Hanselman talk asserting that in an era where anyone can build anything, what gets tested is design skill