Tech • AI • Robotics • Game
Huge Gemini 4 Hydrogen Leak! GPT-7 Bel Breakthrough, Genie 4 DeepSeek V5, GLM 5.5, & Qwen 4! AI News
9.4/10TL;DR
A wave of new AI model sightings, research claims and product launches from Google, OpenAI, Anthropic, Alibaba and others suggests an intensifying race across multimodal systems, cybersecurity, science and on-device AI.
KEY POINTS
Google backend leak points to Gemini 4 Hydrogen
A new model slug, Hydrogen, was reportedly spotted in Google backend logs and is widely suspected to be a Gemini 4 Flash variant. Early reports describe output speeds of about 300 tokens per second, indicating a strong emphasis on low-latency responses. The naming also fits Google’s recent use of element-themed internal codenames.
Genie 4 expands Google’s push into interactive world models
Google has unveiled Genie 4, describing it as its most realistic interactive world model so far. Unlike a standard video generator, it creates explorable environments with real-time prompting, character interaction, native audio and evolving scenes, all built on Gemini Omni. Google is first rolling it out to trusted testers, with access planned through Project Genie, Google AI Studio and Google AI Ultra.
OpenAI model dubbed FE Pichu appears in Arena
A new Arena model called FE Pichu is believed to be tied to OpenAI because it reportedly uses the company’s tokenizer. Its purpose remains unclear, but the appearance has triggered speculation that OpenAI is testing another frontier system ahead of a broader release. No official identification has been confirmed.
Anthropic pushes into drug discovery
Anthropic says its researchers are working with six organizations on new treatments while also building an internal drug-discovery program. In one example, Claude reportedly helped generate about 35,000 antibody designs in a month, coordinated work across 11 external labs, and identified seven manufacturing issues that human reviewers had missed. Anthropic also reportedly acquired Mithro Medicine, a one-person biotech startup that used Claude and external labs to move a rare muscular dystrophy treatment from idea to promising mouse results in under 18 months.
Unreleased Anthropic model reportedly found major vulnerabilities
Anthropic said an unreleased internal model helped uncover a serious Safari same-origin policy bypass. The technique reportedly relied on manipulating limited browser process memory rather than conventional code execution, potentially exposing sensitive data such as email or banking information. The company also said the model helped identify CVE-2026-33096, a Windows HTTP/3 denial-of-service flaw triggered by specially crafted packets.
Math claims raise the stakes in AI-assisted research
Internal models at OpenAI and Anthropic are both said to have produced separate approaches challenging the Exponential Time Hypothesis, a major assumption in theoretical computer science. OpenAI’s work reportedly included arguments around faster exact counting for 3-SAT, while Anthropic said its system generated roughly 14 million output tokens and a formally checked result across more than 1,000 Lean modules. Independent verification remains essential, and none of the claims should be treated as settled.
New model sightings broaden the competitive field
GLM 5.5 has reportedly appeared in Open Code, suggesting testing activity by Zhipu AI rather than a confirmed launch. A new Qwen 4 checkpoint from Alibaba was also spotted, with early demonstrations indicating strong image and video generation quality, including detailed animation and cinematic scene control. Separately, test files reportedly reference DeepSeek V5, V5 Open, and V5 Next, hinting that another release may be close.
Liquid AI targets phones with tiny multimodal models
Liquid AI introduced LFM3, a family of open-weight models designed for local use on phones and other constrained devices. The lineup includes models around 1.7 billion to 2 billion parameters with support for text, vision, audio and structured decision-making, reportedly running with as little as 1 GB of memory. One text model offers a 256K context window, underscoring the growing push toward capable on-device AI.
Tooling and benchmark updates show a maturing ecosystem
Artificial Analysis released Intelligence Index v5, shifting 60% of scoring to private test sets to reduce benchmark contamination. Microsoft introduced MAI Image 2.7 Preview and MAI Cyber 1.1 Flash, while Gemini Notebooks on mobile gained up to one hour of live audio recording as a notebook source. Anthropic is also offering eligible Claude Max users $100 in API credits expiring on November 14.
CONCLUSION
The latest leaks, launches and research claims show the AI race expanding well beyond chatbots into security, biology, mathematics, simulation and edge computing. The central question is no longer whether progress is accelerating, but which companies can turn that progress into reliable products and verified scientific advances first.
Ask a question