Leadership is asking: are we getting value from AI? Which tools are worth the spend? Where are we exposed? Right now, most teams have no idea.

Harmonic Security Usage Explorer changes that. It automatically classifies every AI interaction across your organization into the use cases driving real work, specific to your business. Not generic categories. Not raw prompts. Actual patterns to understand: how your teams are using AI, how much time they spend in AI, the cost, and where risk lives.

CIOs get the data to rationalize spend and cut wasted licenses. CISOs get risk in context. AI committees get proof of impact.

Early access is now open to a limited number of organizations. Request your spot.

Get early access

OpenAI is replacing GPT-5.3 Instant with GPT-5.5 Instant as ChatGPT’s default model. The company says the new version gives clearer, shorter, and more accurate answers while keeping a more natural tone. It is designed for everyday use, meaning small improvements matter because this is the model many people interact with most often.

OpenAI says GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts involving areas like medicine, law, and finance. It also improved on harder conversations that users had flagged for factual mistakes. The model is also better at image analysis, STEM questions, and deciding when web search is needed.

OpenAI is also adding stronger personalization. GPT-5.5 Instant can use past chats, files, and connected Gmail when helpful, while new “memory sources” let users see what context influenced a response and remove outdated information. The model is rolling out to all ChatGPT users, while enhanced personalization starts with Plus and Pro users on web before expanding more broadly.

OpenAI has reportedly raised more than $4 billion from investors including TPG, Brookfield Asset Management, Advent, and Bain Capital for a new company focused on helping businesses use OpenAI’s software. The venture is called The Deployment Company and is valued at $10 billion, not including the money raised. OpenAI is expected to remain the majority owner and controller.

The goal is simple: turn powerful AI models into working business tools. Many companies want AI, but they do not have enough internal engineers to connect models to their data, systems, and daily workflows. This is why OpenAI is leaning into “forward-deployed engineers,” meaning technical teams who work directly with customers to build practical AI systems inside real operations.

The venture’s partners reportedly have access to more than 2,000 portfolio companies and clients, giving OpenAI a huge channel for enterprise adoption. The bigger message is that OpenAI is no longer only competing on model quality. It is now competing on implementation: who can actually get AI working inside businesses fastest.

Anthropic announced a new AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs. The company will focus on helping mid-sized businesses use Claude in important operations, not just as a chatbot. Anthropic says its applied AI engineers will work with the new firm’s engineers to find useful use cases, build custom systems, and support customers over time.

The target customers include community banks, mid-sized manufacturers, regional health systems, and other businesses that could benefit from AI but lack the engineering resources to deploy it properly. Anthropic gives healthcare as an example: Claude-powered systems could help with documentation, medical coding, prior authorizations, and compliance reviews, freeing clinicians to spend more time with patients.

This new company will also join Anthropic’s Claude Partner Network, which already includes major consulting and integration firms like Accenture, Deloitte, and PwC. In plain terms, Anthropic is expanding the number of people who can help companies actually install and use Claude in the messy reality of business operations.

Let ViT Speak: Generative Language-Image Pre-training

ByteDance’s GenLIP trains an image model to describe pictures directly, instead of using extra matching parts. The simpler design worked surprisingly well: with less training data, it matched or beat stronger systems, especially on reading text in images, charts, and documents. The paper argues simpler vision training can scale better.

MolmoAct2: Action Reasoning Models for Real-world Deployment

MolmoAct2 is an open robot brain that connects sight, language, and movement. It adds new robot training data, a better way to understand space, and a faster reasoning mode that focuses only on changed parts of a scene. The authors say it beats rival systems and release the full system.

Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling

This paper tests whether German AI models need more varied text or better text repeated more often. The answer: quality wins. Training on carefully filtered German web pages, even repeated seven times, beat training once on a larger messy set. The team released BOLDT models and cleaned German test sets.

Claude Goes After Wall Street Work Anthropic is pitching Claude as a finance work tool for banks, insurers, asset managers, and fintech companies. It can help build pitch books, credit memos, fund reports, and compliance checks, while letting teams trace numbers back to the source so they can verify the work.

Google Makes Gemma 4 Faster Google released a speed upgrade for Gemma 4 using “multi-token prediction,” meaning the model can guess several next pieces of text at once instead of one at a time. Google says this can make Gemma 4 up to 3x faster without lowering answer quality, which matters for chat apps, coding tools, and phone-based AI.

Anthropic Plans a Massive Google Cloud Spend Anthropic has reportedly committed to spend $200 billion over five years on Google Cloud and Google-linked AI chips. That shows how expensive the AI race has become, because leading models now need huge amounts of computing power, meaning the machines and data centers needed to train and run AI.

Gemini File Search Can Now Read Images Too Google expanded Gemini API File Search so apps can search across both text and images, not just documents. It also added custom labels and page citations, so developers can filter files more easily and show exactly where an answer came from.

Microsoft Turns Copilot Into a Task Worker Microsoft’s Copilot Cowork is moving from answering questions to doing actual work, like inbox tasks, research, documents, and web pages. It is also coming to iOS and Android, with reusable “skills,” meaning saved instructions for repeated tasks, plus plugins that connect it to business tools like Power BI, Dynamics 365, LSEG, Miro, monday.com, and S&P Global Energy.

Big AI Labs Give the U.S. Early Model Access Google, Microsoft, and xAI agreed to let the U.S. government test early versions of new AI models before public release. The goal is to find national security risks, especially cyberattack risks, before powerful AI tools are widely available.

Cerebras Targets a Huge AI Chip IPO Cerebras is aiming for a valuation of up to $26.6 billion in its U.S. IPO, meaning its first sale of shares to public investors. The company makes large AI chips meant to compete with Nvidia, and the offering will test how much public investors still want to bet on AI infrastructure.

OpenAI Wants Musk’s Message Used in Court In a court filing, OpenAI’s lawyers asked to introduce testimony about a message Elon Musk allegedly sent Greg Brockman shortly before trial. They argue the message should be allowed because it may show Musk’s motive and bias in the lawsuit, not because it proves or disproves the legal claims themselves.