Full Summary
This Wednesday morning, two major developments are reshaping the AI landscape: new infrastructure is launching in India, and Google is refining its AI models to reduce token consumption. Both BusinessLine and Express Computer confirm that Neysa and Pipeshift are partnering to launch real-time AI inference infrastructure fully deployed within India. This addresses the current reliance on overseas infrastructure, which leads to unpredictable latency and rising costs for Indian businesses. The new platform, extending Neysa's Velocis system, allows enterprises to deploy open-source models like Gemma, Llama, and Mistral with OpenAI-compatible APIs, keeping all data and inference within India. This is available immediately for use in areas like customer support and workflow automation. Meanwhile, zoomnews.in reports Google has launched Gemini 3.5 Flash Low, specifically designed to reduce token consumption for users of its Antigravity coding platform. Developers previously faced exhausted token quotas, but this new variant uses 45% fewer output tokens without performance compromise, claiming superior results in software engineering tasks. Google has also restructured its Flash model lineup into Low, Medium, and High versions and reset quotas for all plans, including free users. In other AI news, Anthropic's "Mythos" AI model is causing a "Mythos shock" in global security, according to 아시아경제, quickly finding system weaknesses and creating attack plans. Korea's Ministry of Science and ICT is engaging with Anthropic and OpenAI to share vulnerability information, with Anthropic launching "Project Glasswing" to manage Mythos's impact. Geeky Gadgets adds that leaked Anthropic models include Opus 4.8, Sonnet 4.8, and Mythos 1, with Opus and Sonnet bringing efficiency, vision processing, and coding upgrades. Anthropic is also partnering with SpaceX to address computational resource constraints. Open Source For You reveals a TELUS Digital study showing open-source AI models, like GLM 4.7, can match or surpass proprietary models in safety, challenging previous assumptions. ForkLog reports IBM Quantum has enhanced Meta's Llama 3.1 LLM, reducing its perplexity by 1.4% with a 156-qubit processor, a first for end-to-end quantum enhancement in text generation. TechCrunch announces ElevenLabs Music v2 can now switch genres mid-track and handle complex vocals, offering artists new creative tools. Finally, Decrypt highlights a new benchmark, Claw-Anything, where OpenAI's GPT-5.5 scored only 34.5% on complex personal assistant tasks, indicating current AI models struggle with real-world, long-term proactive assistance. And on the longevity front, Yahoo Finance Singapore and BioPharma APAC both report Human Longevity, Inc. and Insilico Medicine are launching a multi-million-dollar collaboration to build AI foundation models for human longevity science, aiming to predict disease risk and accelerate treatments while ensuring data privacy. These developments mean businesses can access more secure, localized, and efficient AI tools, but consumers should be aware that AI personal assistants still have significant limitations in complex tasks.