Anthropic officially launched Computer Use through Claude Cowork and Claude Code on the Mac. This update allows the Claude model to autonomously control Mac environments to execute terminal commands, manage files, and complete complex technical tasks remotely. This transitions Claude from a conversational assistant into an autonomous agent capable of navigating file systems and running apps. The features are currently available to subscribers on the Claude Pro and Max plans.The new Claude Dispatch feature enables interaction with Claude agents from your mobile or web chat interface, allowing for remote control of AI agents (similar to OpenClaw).The bigger picture is that Claude Code and Claude Cowork have been rapidly improving and expanding, with the Claude team shipping 74 features in 52 days. Each feature is incremental, but the sum total is revolutionary. They are creating in Claude Cowork an AI digital worker capable of executing complex, multi-step tasks across any software on a user’s machine. Feel the AI acceleration.Google released Gemini 3.1 Flash Live, a low-latency multi-modal model optimized for real-time voice and video interaction; it can process text, images, audio, and video and return immediate spoken responses. Google’s model card shows Gemini 3.1 Flash Live is their highest-quality audio-to-audio and voice model yet, beating out Grok Voice Agent and prior Gemini models and scoring 95.9% on BigBench audio. It is available via API, Gemini app, and the expanded Search Live experience. Mistral released Voxtral TTS, an open-weight multilingual TTS model with low latency and support for emotionally expressive speech in nine languages. Voxtral TTS is lightweight enough for scalable local deployment and matches well against ElevenLabs Flash in human preference tests. Voxtral TTS can be downloaded from HuggingFace and run on local servers, which opens up data-sensitive enterprise use cases and local applications.Unsloth Studio was updated to offer 10x faster local interface, desktop shortcuts, and auto-parameter detection. Unsloth Studio is a web UI for training and running models locally across Windows, Linux, and macOS, with support for chat, training, exports, and multimodal files.Reka AI launched Reka Edge, an open 7B multimodal vision-language model built for sub-second latency on edge devices. Reka Edge accepts image or video plus text input and is optimized for image understanding, video analysis, object detection, and agentic tool use. It’s now available on OpenRouter and Hugging Face.Modular announced Mojo 26.2 can run FLUX.2 image generation in under a second and at dramatically lower cost. Mojo is a high-performance Python-like language that supports GPU optimizations in AI stacks. This enables a “fundamentally different cost structure for image generation at scale” and can accelerate other AI workloads.Figma has launched its AI agents into beta, introducing Figma Model Context Protocol (MCP) server via the use_figma tool. These agents can design directly on a live Figma canvas while maintaining a full understanding of the user’s existing design system and context. The tool is designed to assist teams by automating UI components and layout tasks with high precision.Luma Labs released Uni-1, an AI image generation model that “thinks and generates pixels simultaneously” and supports iterative interaction while creating images. Uni-1 is intelligent and directable, supports chat-style collaboration in editing images, and can also generate infographics. Luma Labs claims it is frontier-levels, with Arena ELO scores on par with Nano Banana 2.Cohere launched Transcribe, an open-source 2B automatic speech recognition model (audio-in, text out) with a best-in-class 5.42% word error rate as shown by the Hugging Face open ASR leaderboard page. Transcribe supports 14 languages and can be deployed in local and enterprise-friendly speech recognition applications.Google DeepMind released Lyria 3, their most advanced music generation model yet. Lyria 3 can generate full three-minute music tracks with structural control over intros, verses, choruses, and bridges. It can compose from images, and it applies SynthID watermarking. This makes it a significant generative-audio upgrade that gets AI closer to production-ready music workflows.OpenAI introduced a new suite of safety tools for younger users, including an open-source teen safety classifier called gpt-oss-safeguard-20b. The company also updated its Model Spec with “Under-18 Principles” and launched new parental controls for ChatGPT.Phota Labs launched Phota Studio and Phota API, a photography-focused image generation and editing model with identity-preserving personalization. The Phota specialized image model provides personalized photo generation from a user’s uploaded images with stronger identity consistency than generic image models usually provide. Figure 3. Phota can remove that ex- from your photos while keeping the rest of your photos accurate and clean.Just-released Irodori-TTS-500M is an open Japanese language TTS that includes a distinctive emoji-based emotion control. It supports zero-shot voice cloning and lets users steer style, emotion, and even sound effects by inserting emojis into the prompt text.An upcoming powerful Anthropic model called Claude Mythos was confirmed after a data leak on their website. The rumored 10T parameter model offers a “step change” in AI performance and would occupy a new tier above the current Opus flagship. Documents reveal significant cybersecurity risks associated with the powerful model. Google Research introduced TurboQuant, a compression technique that reduces AI model KV (key-value) cache memory by more than 6-fold while maintaining AI model performance. Presented in “TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate,” TurboQuant compresses vector data into more compact polar coordinates, that with just 3 bits of quantization preserve full accuracy of 16 bit KV cache memory. This allows for an 8-fold speedup in long-context model inference and vector search with zero loss in accuracy, significantly improving the efficiency of AI models and allowing for much faster, cheaper AI inference.ARC Prize launched their new AI intelligence benchmark ARC-AGI-3, an “unbeaten benchmark” where humans score 100% and today’s frontier AI systems score under 1%. As shared in their technical paper, ARC-AGI-3 measures agentic intelligence via human-like generalization on novel tasks.Figure 4. The AI progress has accelerated so much that ARC-AGI-1 was saturated in 2025, and ARC-AGI-2 is being saturated in 2026. Time for a new benchmark, ARC-AGI-3.The Allen Institute (AI2) has released MolmoWeb, a new open visual web agent designed to navigate the web by interpreting screenshots. Built on Molmo 2 4B and 8B parameter multi-modal models, MolmoWeb agents navigate websites using only raw screenshots, and they outperformsGPT-4o on major web navigation benchmarks. The release is on HuggingFace and includes the MolmoWebMix dataset, consisting of over 150,000 interaction trajectories.OpenAI has officially shuttered its Sora video generation platform and API just six months after its wide release, as the company refocused on productivity and agents. This decision terminates a high-profile $1 billion partnership with Disney, which included a licensing agreement for over 200 characters. The shutdown is reportedly due to the extreme GPU resource consumption required to maintain Sora as a service. OpenAI also canceled their “adult mode” ChatGPT and ChatGPT Instant Checkout as part of a strategic shift back toward core products.In a major leadership restructuring, OpenAI CEO Sam Altman is relinquishing direct oversight of the company’s safety and security teams to focus on infrastructure scaling. Altman’s priority has shifted toward managing buildouts of data centers and capital raising; he recently secured an additional $10 billion in funding for a $120 billion round. Concurrently, OpenAI completed initial development of its next-generation model, codenamed “Spud.”The OpenAI Foundation has been formally announced. Led by Bret Taylor, the organization holds approximately $130 billion in equity in OpenAI to support AI-driven scientific discovery, including work on curing diseases, and societal resilience due to AI. The organization has committed over $1 billion toward research in these areas. This completes the process where OpenAI the company and OpenAI the non-profit foundation re-organized to become distinct entities.Facing massive compute demand from users running Claude tools, Anthropic has reduced daily rate limits on its platforms, stating that during weekday peak hours users would use up five-hour session limits faster while weekly limits remained unchanged. Anthropic’s help center also offered a promotion that doubled usage outside peak windows, while Claude users reported burning through Pro and Max quotas much faster than before.Alphabet’s Google DeepMind has partnered with Munich-based Agile Robots to deploy its Gemini Robotics foundation models across their global fleet of industrial robotic hardware. By bringing advanced reasoning capabilities directly into real-world factory environments, the collaboration uses physical world feedback to rapidly improve the models and the robots’ industrial adaptability.Robotics funding has hit mega-round territory. Mind Robotics ($500M), Rhoda AI ($450M), Sunday ($165M, now a unicorn), and Oxa ($103M) have collectively raised over $1.2 billion in recent weeks, indicating that humanoid and autonomous robotics has entered the same mega-round levels that defined AI labs and LLM makers in recent years.Amazon has entered the consumer humanoid robotics market by acquiring Fauna Robotics, a New York-based startup. Fauna is the creator of “Sprout,” a lightweight, bipedal robot designed for human-centric spaces like homes and schools.AI-powered notetaker app maker Granola raised $125M at a $1.5B valuation.A new AI startup called Hark Labs launched with $100 million in seed funding, led by a former Apple design chief to build custom hardware-model integrations.In Anthropic’s lawsuit over the U.S. Department of War’s supply-chain-risk designation, Federal U.S. District Judge Rita Lin temporarily blocked the Pentagon from branding Anthropic a supply chain risk, pending further proceedings. The Judge’s order held that the department had not met the legal requirements for that action.The city of Baltimore filed a lawsuit against xAI over the generation of deepfake pornography. The lawsuit alleges that the Grok chatbot lacks the necessary guardrails to prevent users from creating non-consensual sexual imagery. The case could set a legal precedent for how cities hold AI developers accountable for harmful model outputs.Progressive lawmakers Bernie Sanders and AOC have proposed a bill that would place restrictions on new data center development, citing energy grid strain and environmental concerns. Meanwhile, new U.S. data center additions halved in Q4 2025 due to grid bottlenecks and power queue delays, signaling a slowdown in infrastructure expansion.The March 2026 Anthropic Economic Index shows that long-term Claude users have a 10% higher success rate in completing complex tasks than new users. “AI literacy” aids productivity, as experienced users (users with six or more months of Claude use) iterate more effectively. Additionally, the report shows professional coding is rapidly migrating to API-based workflows via Claude Code, away from the chat interface.