Figure 1. xAI released Grok Imagine Video 1.5 to general availability. It can generate high quality 720p audio-video clips in under a minute.Z.ai launched GLM-5.2, a 753-billion parameter Mixture-of-Experts (MoE) open-weights AI model for long-horizon coding and engineering tasks. GLM-5.2 features a 1-million-token context window, reasoning controls, and support for coding tasks across entire codebases; it also utilizes the IndexShare architecture to reduce per-token compute FLOPs by up to 2.9 times. GLM-5.2 demonstrates high performance on benchmarks like SWE-bench Pro (62.1%), Terminal-Bench 2.1 (81.0%), and FrontierSWE (74.4%), rivaling frontier models like GPT-5.5 and Claude Opus 4.8.Independent evaluations by Artificial Analysis confirm GLM-5.2 as the leading open weights AI model on the Artificial Analysis Intelligence Index. It also shows GLM-5.2 is notably token-hungry, consuming roughly 43,000 output tokens per standard Index task, up from 26,000 tokens used by GLM-5.1. GLM-5.2 is priced at only $1.40/$0.26/$4.40 per 1M input/cache hit/output tokens, so despite the token-hungry reasoning, this is the lowest-cost frontier-level AI model, substantially cheaper than proprietary rivals.Figure 2. GLM-5.2 shows superior performance to GLM-5.1 and even Claude Opus 4.7 at High effort. Greater reasoning effort with more token use leads to higher performance.GLM-5.2 is positioned as a powerful open model focused on agentic software engineering that developers can run and build on. The GLM-5.2 model is available through Z.ai’s Coding Plan, their ZCode agent, their Z.ai chatbot, and via open weights on HuggingFace.Sina Weibo researchers released VibeThinker-3B, a 3B parameter model matching flagship reasoning performance. The model achieves a score of 94.3 on AIME 2026 and 80.2 LiveCodeBench v6, a remarkable out-performance for a 3B model that matches top-tier AI models such as Claude Opus 4.5. This has stimulated conversation about benchmark and scaling limits.The Technical Report on VibeThinker-3B shows that verifiable reasoning capabilities can be compressed into much smaller models than those required for open-domain knowledge, calling this the Parametric Compression-Coverage Hypothesis. This could have huge implications for how much further we can compress AI reasoning and improve AI model efficiency.xAI released Grok Imagine Video 1.5 to general availability, xAI’s image-to-video system for generating short clips with synchronized sound. The update features improved motion physics, better audio synchronization (same-pass audio and speech generation), and nearly doubled generation speeds for 720p videos; its fast mode produces 6-second 720p videos in about 25 seconds. The release also adds workflow features such as Projects, multiple parallel agents, and search. Results are compelling:You type a prompt or upload an image, and it turns it into a realistic 720p video, up to 15 seconds long, with actual dialogue and sound effects. All in just 25 seconds.Google made Gemini Omni available through an API and positioned it as a leading video model. Gemini Omni is Google’s unified any-to-any system for text, image, video, audio, and music generation and editing. Google’s model page says Omni performs strongly on video editing, text-to-video, image-to-video, and reference-to-video, and reports top results on MovieGenBench for overall preference and instruction following. The model is meant for iterative multimodal video creation, including continuation, reference-based edits, and consistency across turns.Anthropic has overhauled Claude Design, introducing enhanced canvas controls for easier element manipulation and brand-compliant design system imports from GitHub or local files. The update expands integration and export capabilities to platforms like Adobe, Canva, and Vercel, while also implementing shared usage limits across Anthropic’s product suite. Additionally, a new `/design-sync` command enables seamless, bidirectional workflow synchronization between Claude Design and Claude Code.Nvidia released XR AI in public beta, a developer framework for building multimodal AI agents that run on AR glasses and extended-reality devices. The system connects video, audio, depth, pose, and sensor data with enterprise retrieval, AI models, agent orchestration, and accelerated inference. This enables hands-free AI assistance in laboratories, factories, hospitals, and design workflows.HumanLayer launched its Agentic IDE for teams working in complex codebases. Declaring they are on a mission to ‘solve the AI slop code problem’, HumanLayer is aiming their HumanLayer Agentic IDE at structured, team-based AI software development rather than one-shot vibe coding. It includes a collaboration platform and software-factory building blocks designed to help engineers ship 3x faster while maintaining code quality and standards.OpenRouter introduced Fusion, which combines access to multiple AI models behind one API call. Fusion sends one prompt to a panel of models, has a judge model compare the outputs, and then synthesizes a final answer from the combined results. In OpenRouter’s reported DRACO tests, fused panels outperformed individual models, and a lower-cost panel came within about 1 percentage point of Fable 5 while costing about half as much.Midjourney announced Midjourney Medical, a new business to build a full-body ultrasound scanner called Ultrasonic CT that can do whole-body ultrasound scans in as little as 60 seconds, then offer it as a service in Midjourney Medical spas. The ultrasound system uses thousands of ultrasonic transducers to build a 3D anatomical map. It seems like a big leap to go from image generation into medical imaging hardware and services, but Midjourney Medical notes that large data volumes would make AI useful for processing and reconstruction.Figure 4. Midjourney Medical’s ultrasound system can do whole-body scans and map a number of anatomical features and medical conditions.Samsung announced a new AI-powered pet health feature for mobile devices during the VivaTech 2026 conference in Paris. Developed in collaboration with the platform Lifet, the feature uses AI to analyze photos of pets to detect conditions such as obesity and periodontal disease.Tokyo-based AI startup Sakana AI has launched its first commercial product, Sakana Marlin, an autonomous AI research agent that works for up to eight hours to deliver deeply researched 100-page strategy reports and executive slides. The Sakana Marlin platform is designed exclusively for enterprise use and features a strict data policy ensuring customer inputs are never used for model training without consent.OpenAI updated its platform deprecation notices for older GPT-5 and o3 model snapshots. Older GPT-5 and o3 snapshots will be removed from the API on December 11, 2026, while older GPT Image models and other legacy model families also have scheduled removals.Researchers from multiple US Universities released SciAgentArena, a benchmark for evaluating AI agents in realistic scientific research scenarios. The benchmark includes roughly 200 tasks with stepwise verification, and an interactive agent-agnostic environment to assess AI agents. Benchmark results show that current agents are useful for well-specified data-analysis workflows but weaker at novel insight generation, exploration, and robust open-ended scientific reasoning.Google DeepMind published “From AGI to ASI,” which explains Artificial Superintelligence (ASI) as systems surpassing large human organizations in capability and investigates the transition to improving AI from AGI to ASI. It explores four potential development pathways: scaling, paradigm shifts, recursive self-improvement, and multi-agent collectives. Each path has bottlenecks, and AI progress may accelerate continuously rather than in a single step change. Their roadmap shows ASI is attainable in the near future, requiring a global effort to prepare for coming transformative societal shifts.SpaceX bought Cursor’s parent company, Anysphere, in a $60 billion stock deal. The acquisition followed the massive SpaceX IPO and consolidates the AI landscape, bringing Cursor’s AI coding application and data into xAI’s broader AI model and AI infrastructure efforts. This acquisition integrates Cursor’s user base into SpaceX’s AI unit while bolstering Cursor’s position, as Cursor’s market share among AI coding tools slipped to 26% amid intense competition from tools like Claude Code.Enterprise software ecosystems are undergoing an aggressive shift in pricing structures as CIOs push back against traditional seat-based subscription models in favor of consumption-based or outcome-focused metrics. Because autonomous AI agents operate independently of human headcount, software vendors are rewriting their commercial terms to charge based on API token volume, compute utilization, or verified task completion.DeepSeek raised more than $7.4 billion in a funding round that valued the company at more than $50 billion, making it the most valuable Chinese AI startup. Its founder, Liang Wenfeng, invested around $3 billion in the fundraise. He previously held nearly 90% of the company before the financing round. A government-backed fund invested around $150 million.At the G7 meeting, French President Emmanuel Macron urged the U.S. to share cutting-edge AI and called for democratic cooperation on regulation, in the wake of U.S. restrictions on Anthropic’s Fable 5 and Mythos 5 models. Macron criticized unilateral restrictions on Anthropic’s models as too nationalist.Likewise, European Commission President Ursula von der Leyen said it is in both U.S. and EU interests for Europe to have access to the best AI models. The EU wants shared access to frontier AI capabilities under common safety standards rather than a drift toward nationalist AI controls.OpenAI CEO Sam Altman and other AI leaders supported an international coalition for AI safety standards at an AI CEOs and leaders meeting at the G7 summit, with AI tech leaders proposing international cooperation with democratic oversight over AI deployments.Anthropic and Tata Consultancy Services announced a partnership to bring Claude to regulated industries. TCS will provide Claude to 50,000 employees in 56 countries, build Claude-powered products for financial services, healthcare, public-sector, aviation, telecom, and life-sciences clients, and join the Claude Partner Network.Anthropic also announced a multi-year global alliance with DXC Technology. DXC will train tens of thousands of Claude-certified forward-deployed engineers and integrate Claude into systems used by banks, airlines, insurers, manufacturers, and government agencies. DXC notes that Claude was used to generate more than 95% of the code for DXC OASIS, its AI-native managed-services orchestration platform.Jeff Bezos argued AI will ultimately create labor shortages rather than mass unemployment. Speaking at VivaTech, Bezos framed AI as a productivity accelerator that will expand the economy and create new kinds of work, a sharply optimistic contrast to surveys showing widespread job-loss concern. His argument captures executive optimism about how AI is a door to more opportunities rather than thinking of the economic possibilities as static.“I promise you every single person in this audience has had an idea for a new business or a new product or a new device that they wish they could manufacture, and that idea stayed in your head and went nowhere. And the reason it stayed in your head and went nowhere is because it’s too hard to do, and it wasn’t worth it.If we can accelerate the dream build loop, all of the ideas will then become possible. And then we end up being limited not by our capabilities, but by our imaginations. – Jeff Bezos
AI Week in Review 26.06.19
Grok Imagine 1.5, GLM-5.2, VibeThinker-3B, Claude Design re-designed, Nvidia XR AI, HumanLayer Agentic IDE, OpenRouter Fusion, Midjourney Medical, Sakana Marlin, SciAgentArena, DeepMind's AGI to ASI.










