Google Cuts "AI Plus" Price by 40% — Now ¥725/Month, with Storage Doubled

Google has announced a major pricing overhaul for its personal AI subscription, "Google AI Plus." The reduction from ¥1,200 to ¥725 per month amounts to a roughly 40% price cut, and storage capacity is being doubled at the same time.

According to ITmedia, Google has already implemented price cuts and added perks for its higher-tier "AI Pro" and "AI Ultra" plans, so this Plus revision completes a full round of enhancements across all three tiers. In an AI services market crowded with rivals like ChatGPT Plus ($20/month) and Anthropic's Claude Pro, Google is aiming to win market share on price.

The move suggests that AI services are becoming increasingly commoditized. As the performance gap between models narrows, we are entering an era where pricing directly determines subscriber acquisition and retention.

JetBrains' "Mellum 2" Earns High Praise in the Coding Model Scene

Mellum 2, a coding-focused language model released by JetBrains, is drawing attention in Reddit's LocalLLaMA community. The model has been described as "really good and performant," and it reportedly delivers strong performance on code completion and generation tasks.

JetBrains is best known for IDEs such as IntelliJ IDEA and PyCharm, and by offering its own in-house AI models, the company is bringing the AI-assisted development experience under its own roof. Coding-specialized models tend to deliver practical performance even at smaller sizes, which also makes them attractive from a local-execution standpoint.

AI Predicts 10-Year Stroke Risk from Routine Cardiac Exams

A research team at Mass General Brigham has announced the development of an AI model that can predict a patient's stroke risk 10 years into the future from routine cardiac exam data.

Stroke is a condition where early detection and prevention are critically important, and conventional risk assessment methods have their limits. If AI models enable more precise long-term forecasting, they could significantly support decision-making around preventive treatment. For now, the work remains at the research stage, and further validation is expected to be required before clinical deployment.

New AI Spy Powers Spark Putin's "Camera Concerns"

According to the Financial Times, the expansion of new AI-driven intelligence powers has raised concerns for Russian President Putin over "surveillance via cameras."

Advances in AI-based image recognition and surveillance technology are reshaping how nations conduct espionage. The report shows that AI's impact on the national security domain has become a reality. The AI-ification of surveillance technology is also igniting debate over the balance with privacy and civil liberties.

llama.cpp's WebGPU Gets Faster k-quants — A Boost for Practical Browser Inference

A Pull Request that improves prefill speed for k-quant models in llama.cpp's WebGPU backend is making waves on Reddit. A matmul refactoring targeting Q4/Q5/Q8 quantization and k-quants is expected to improve LLM inference performance in the browser.

Efforts to run LLMs in the browser via WebGPU expand the possibility of using AI without installing a dedicated app. Faster prefill (prompt processing) is a key improvement that directly translates into a better user experience.

Parakeet 0.6B Fine-Tuned for Medical ASR — Open Weights Released

Reddit's LocalLLaMA community has reported a case of fine-tuning NVIDIA's speech recognition model "Parakeet 0.6B" for medical-domain speech recognition (Medical ASR). The weights have been released openly, and the model reportedly runs locally on Mac/CUDA/CPU.

While demand for voice documentation in medical settings is high, the use of cloud APIs is constrained by the need to protect patient data privacy. A locally runnable, medical-specialized ASR model stands out as a promising approach to this challenge.