AI news — July 21, 2026
The 20 top stories PulseAugur surfaced that day, ranked by signal across labs, papers, and developer communities.
-
Google launches Gemini 3.6 Flash and 3.5 Flash-Lite for enterprise AI agents
Google has launched two new Gemini models, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, aimed at reducing latency and token costs for enterprise AI agents. Gemini 3.6 Flash is designed for tasks requiring coding and multimodal reasoning, offering improved performance and reduced …
-
Google launches cost-effective AI security model Gemini 3.5 Flash Cyber
Google has introduced Gemini 3.5 Flash Cyber, a new AI model designed for cybersecurity tasks, aiming to be a more cost-effective option compared to larger, more expensive models like Anthropic's Mythos. This specialized model integrates with Google's CodeMender tool to rapidly …
-
Jushi Intelligence unveils world's first embodied semantic intelligence system
Jushi Intelligence has launched insightOS Semantic, the world's first embodied semantic intelligence system. This system has been validated in real-world scenarios such as warehouse picking and delivery, disassembly and palletizing, multi-machine collaborative pipelines, and sch…
-
Grok AI accessible in Russia, latest model Grok 4.5 released
Grok, the AI chatbot from Elon Musk's xAI, is accessible in Russia despite the broader X platform being blocked. Users can access grok.com directly, though unstable IP access may require a VPN. The latest flagship model, Grok 4.5, released on July 8, 2026, is described as an "Op…
-
K3 AI model launch causes server strain and market shifts
K3, a new AI model, has seen overwhelming demand since its release, causing server overloads and a significant drop in NVIDIA's stock price. Anthropic has reportedly adjusted its subscription services in response to the surge in interest and competition. The rapid adoption and i…
-
OpenAI launches GPT-Live-1, enabling simultaneous listening and speaking
OpenAI has released GPT-Live-1 and GPT-Live-1 mini, new full-duplex voice AI models that can listen and speak simultaneously, mimicking human conversation flow. These models replace the previous Advanced Voice Mode and are rolling out globally across various platforms, including…
-
DeepSeek V4 "Full-Power" Version Set for Release Amidst Lab Fraud Claims
A "full-power" version of the DeepSeek V4 model is reportedly set to be released soon, potentially as early as tomorrow. This announcement follows claims that a "mysterious laboratory" which previously topped global rankings was exposed as fraudulent. The news comes amidst signi…
-
Moonshot AI's Kimi K2.7 Code integrated into GitHub Copilot · 1 source tracked
Moonshot AI's Kimi K2.7 Code, an open-weight model with 1 trillion parameters and a 256K context window, has been integrated into GitHub Copilot. This integration occurred remarkably quickly, with the model's weights released on June 12, 2026, and its addition to Copilot's model…
-
DeepSeek V4 'Full Version' Release Imminent Amidst Lab Fraud Exposure
A mysterious lab that claimed to have developed the groundbreaking DeepSeek V4 model has been exposed as a fraud. The actual DeepSeek V4, however, is reportedly set to be released soon, with its full capabilities expected to be unveiled imminently. This revelation follows a peri…
-
Anthropic's Claude 3.5 Sonnet enhances coding, while GPT-4o faces multimodal input issues
Anthropic has released Claude 3.5 Sonnet, a new coding-focused LLM that boasts a 49.0% improvement on the SWE-bench Verified benchmark. This model is designed to reduce hallucinations and enhance developer workflows through a rapid fine-tuning process and efficient context manag…
-
Anthropic IPO gears up with investor meetings, eyes $965B valuation
Anthropic is reportedly preparing for an Initial Public Offering (IPO) with major banks Goldman Sachs, Morgan Stanley, and JPMorgan Chase leading the process. Investor meetings are scheduled to begin on July 15, 2026, with a potential IPO debut in October 2026. The company's val…
-
Google releases new Gemini 3.5 Flash and Lite models
Google has announced the release of its latest Gemini models, including Gemini 3.5 Flash and Gemini 3.5 Lite. These new models are designed to offer enhanced performance and efficiency for a wide range of applications.
-
AssemblyAI explains advanced AI voice recognition capabilities
AssemblyAI has detailed how modern AI voice recognition technology goes beyond simple speech-to-text conversion. The technology now incorporates machine learning models capable of identifying speakers, detecting sentiment, and analyzing the core topics and themes within a conver…
-
AssemblyAI explains how context improves AI speaker labeling
AssemblyAI has detailed how context, specifically spoken names, influences automatic speaker labeling in audio transcriptions. The process, known as speaker diarization, partitions audio by speaker and assigns consistent labels like 'Speaker A' and 'Speaker B'. Mapping these lab…
-
AssemblyAI details advanced speech recognition evaluation beyond WER
AssemblyAI has published a guide detailing advanced methods for evaluating speech-to-text models, moving beyond the traditional Word Error Rate (WER). The article highlights the limitations of WER, such as its inability to account for semantic meaning or errors in ground truth t…
-
Moonshot AI releases Kimi K3, leading China's open-weight model wave
Moonshot AI has released Kimi K3, a large open-weight model with a 1 million token context window, positioning it competitively against top closed-weight models like Claude Opus 4.8 and GPT-5.5. This release is part of a broader trend of advanced open-weight models emerging from…
-
RoboSense launches E2 perception platform for physical AI data
RoboSense has launched its second-generation all-solid-state perception platform, E2, designed to serve as a data input for physical AI. This new platform utilizes the company's proprietary "Peacock" SPAD-SoC chip and a 2D VCSEL chip, offering significantly improved accuracy and…
-
Google develops 'Rigel' chip to boost Gemini AI efficiency
Google is reportedly developing a proprietary AI chip, codenamed "Rigel," designed to significantly outperform its existing Tensor Processing Units (TPUs) in terms of energy efficiency for its Gemini models. This move suggests a strategic effort by Google to gain a competitive e…
-
HALO v2.7 simplifies AI security agent architecture with unified tool registry
HALO, an open-source AI-powered security agent, has undergone a significant refactor with version 2.7, focusing on architectural simplification and reliability. The update introduces a centralized tool registry featuring 29 integrated security tools, a transport-agnostic executi…
-
Anthropic's Claude Sonnet 5 offers near-Opus quality at lower cost
Anthropic has released Claude Sonnet 5, positioning it as their most agentic Sonnet model yet, capable of planning, tool use, and autonomous task completion that previously required larger models like Opus 4.8. While initially priced significantly lower than Opus, the model's co…