AI Digest

Anthropic Releases Claude 5 with On-Device Reasoning
Anthropic launched Claude 5, its flagship model, featuring a new “local reasoning” mode that allows complex inference to run entirely on a consumer laptop without cloud connectivity. Early benchmarks show a 40% improvement in mathematical reasoning over Claude 4, while the on-device capability marks a significant step toward privacy-preserving AI assistants. The model is available immediately for Pro and Team subscribers.

Google Unveils Gemini 2.5 Pro with Million-Token Context Window
Google DeepMind released Gemini 2.5 Pro, boasting a one-million-token context window—enough to process entire codebases or multi-hour video feeds in a single query. The model also introduces native tool-use APIs for web browsing and code execution, positioning it as a direct competitor to OpenAI’s GPT-5 for enterprise workflows. Pricing remains unchanged from the previous generation.

EU Parliament Passes AI Liability Directive with Strict Transparency Rules
The European Parliament voted overwhelmingly to approve the AI Liability Directive, requiring companies deploying high-risk AI systems to maintain auditable logs of all training data and decision-making pathways. The law imposes fines of up to 7% of global revenue for non-compliance, and will take effect in March 2027. Tech industry groups have already signaled plans to challenge the regulation in court.

Apple’s Mixed-Reality Headset Sales Pass 5 Million Units
Apple announced that its Vision Pro 2 headset has sold over five million units since its January launch, driven by a price drop to $2,499 and the introduction of a lightweight “Air” model. The device now accounts for 18% of Apple’s wearables revenue, with developers reporting a 200% increase in spatial computing app submissions over the past quarter.

OpenAI Quietly Launches GPT-5 Turbo for Developers
OpenAI released GPT-5 Turbo, a faster and cheaper variant of its flagship model, with a 50% reduction in per-token cost and support for up to 256,000 tokens of output. The model is optimized for real-time applications like customer service and coding assistants, but notably lacks the multimodal vision capabilities of the full GPT-5. Developers can access it via API starting today.

Nvidia Announces Blackwell Ultra GPU for Data Centers
Nvidia unveiled the Blackwell Ultra GPU, offering 2.5x the AI training performance of its predecessor while consuming 30% less power per teraflop. The chip integrates 288GB of HBM4 memory and is designed specifically for training trillion-parameter models. Major cloud providers including AWS and Microsoft Azure have already placed pre-orders for Q3 delivery.

China’s ByteDance Releases Open-Source Video Generation Model
ByteDance open-sourced its “DouYing 2.0” video generation model, capable of producing 4K-resolution, 60-second clips from text prompts with consistent character animation. The model, released under a permissive license, has already been downloaded over 100,000 times on Hugging Face. Analysts note it rivals OpenAI’s Sora in quality while being freely available for commercial use.

US Federal Trade Commission Files Antitrust Suit Against Meta Over AI Acquisitions
The FTC filed a lawsuit seeking to unwind Meta’s acquisitions of two AI startups, alleging the company systematically acquired potential competitors to maintain dominance in the large language model space. The complaint specifically cites Meta’s 2023 purchase of a leading open-source model developer and a 2024 deal for a synthetic data company. Meta called the suit “baseless” and vowed to fight it in court.

Editor’s take: Today’s news reveals an industry accelerating on two parallel tracks: unprecedented model capability leaps (Claude 5, Gemini 2.5 Pro) and intensifying regulatory pushback (EU directive, FTC suit). The tension between open-source momentum (ByteDance) and closed-source dominance (OpenAI, Anthropic) is no longer a debate—it is the central structural dynamic shaping the next decade of AI.