michael.nunez@venturebeat.com (Michael Nuñez)

Perplexity launched hybrid compute for its Computer platform, allowing a single AI agent to split work between cloud-based frontier models and locally-running open-weight models on Apple silicon Macs. Sensitive data is routed to the local machine via a trained PII classifier called a Privacy Gate, ensuring confidential information never leaves the device while the agent maintains task context. The feature is available today for enterprise customers and Pro/Max subscribers on macOS 15 or later.

Salesforce and Anthropic announced Claudeforce, a partnership that embeds Salesforce's CRM directly into Claude through a CoWork plugin with 37 pre-built sales skills. The plugin lets users query, update, and act on live CRM data without opening Salesforce's interface, with pilot availability now and open beta in September. The move reflects a strategic bet that enterprise software's future may not require traditional application interfaces.

IBM announced a dual-architecture mainframe processor at Hot Chips that can natively execute both Arm and IBM Z instruction sets on the same cores, switching between them in nanoseconds. Built on 2-nanometer process technology with 11 cores running above 5.7 GHz, the chip allows enterprises to run Arm-native Linux software and AI frameworks alongside z/OS transaction-processing workloads on shared silicon. The design represents the first hardware outcome of IBM and Arm's April strategic collaboration and directly addresses whether mainframes can remain relevant in an AI-dominated infrastructure landscape.

Anthropic updated Claude Tag, its Slack agent, to read full conversation context rather than evaluating messages individually, improving its ability to decide when to interject unprompted by roughly 30%. The company frames this as a shift from single-user AI tools to 'multiplayer AI' agents that operate across teams and pursue organizational goals. Scott White, Anthropic's head of enterprise product, argues this represents AI transitioning from personal assistant to organizational colleague.

Cursor launched Origin, its own code hosting platform, on Monday, coinciding with a six-hour-plus GitHub outage that affected pull requests, issues, APIs, and Copilot. Origin integrates AI agents directly into the code review workflow and maintains GitHub as the source of truth rather than requiring migration. The timing dramatized Cursor's core argument that AI has made code hosting infrastructure a meaningful procurement decision again.

Skan AI raised $63 million in Series C funding to build what it calls a 'context graph of work' by observing how employees actually perform jobs across enterprise software. The company is launching two new products, Skan AI Blueprint and Skan AI Agents, alongside its existing intelligence offering. The funding and product announcements address a widespread problem in enterprise AI, where only 8% of companies have AI agents in production and 95% of early implementations require complete redesign.

Mistral AI announced a three-part infrastructure expansion targeting 1 gigawatt of European compute capacity by 2030, starting from less than 200 megawatts today. The plan includes regional inference endpoints, a new Priority Tier with uptime guarantees, and a coalition of European enterprises making multi-year compute commitments to fund 200 megawatts by end of 2027. The company is also hosting third-party open models, including GLM-5.2 from Chinese AI lab Z.ai, marking a shift from open-weight model training toward critical infrastructure services.

AWS announced at Black Hat USA 2026 that its Continuum vulnerability platform will integrate directly into Anthropic's Claude Code and OpenAI's Codex, embedding AWS security tooling at the point where developers write code regardless of which AI model they use. The move positions AWS as a security control plane for enterprise software development and reflects an urgent industry response to frontier AI models like Claude Mythos Preview, which identified thousands of previously unknown zero-day vulnerabilities during testing. AWS also expanded its Security Hub Extended marketplace with a 10th category focused on supply chain protection, adding Chainguard and Socket as partners.

The Model Context Protocol, an open standard connecting AI agents to enterprise software, released its largest update since launch twenty months ago. The revision transitions MCP to a fully stateless architecture, removes the need for persistent session management, and graduates interactive interfaces and long-running tasks into official protocol extensions. The changes eliminate operational barriers that previously made large-scale production deployments complex, allowing organizations to run MCP servers behind standard load balancers using existing cloud-native tooling.

Snowflake launched Cortex AI Gateway, a centralized control layer for governing how AI agents access enterprise data and tools, alongside security integrations with 1Password, Aembit, Linx Security, SailPoint, and Saviynt. The platform addresses a fundamental security gap: traditional enterprise security assumes humans are the actors, but AI agents operating at machine speed can exploit permission gaps and amplify existing risks. Snowflake positions itself as the control plane that decides what agents can do with enterprise data, rather than allowing each vendor to build closed ecosystems.

Microsoft unveiled MAI-Cyber-1-Flash, a custom-built AI security model, and Project Perception, an agentic defense platform designed to automate vulnerability detection and remediation. The system scores 96% on the CyberGym benchmark while cutting costs roughly in half compared to Microsoft's current production setup. The architecture routes 90% of security tasks to the smaller, cheaper model and escalates the remaining 10% to OpenAI's GPT-5.4, reflecting Microsoft's strategy to compete on cost efficiency rather than raw model size.

Microsoft released two new in-house AI models, MAI-Image-2.5-Pro and MAI-Voice-2-Flash, into public preview, claiming production deployments across its product suite show GPU cost reductions of up to 89% compared with OpenAI models. The announcement represents Microsoft's most concrete argument yet that it can power enterprise products with proprietary models rather than relying on third-party frontier models. Both models are now running in production across Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure.

Anthropic released Claude Opus 5 on Friday, positioning it as a cost-efficient alternative to its flagship Fable 5 model at half the price. The model scores higher than Fable 5 on several coding and agentic benchmarks while maintaining the same token pricing as its predecessor, Opus 4.8. The launch reflects a shift in the AI industry from raw capability competition toward economic efficiency for enterprise workflows.

Poolside, a San Francisco AI lab, released Laguna S 2.1, a 118-billion-parameter coding model that matches or beats open models several times its size on benchmark tasks despite being significantly smaller. The model was trained in under nine weeks on 4,096 Nvidia H200 GPUs and is available immediately under an open license. The release represents a strategic bet that Western labs can compete in open-weight AI through efficiency and iteration speed rather than raw scale, and targets enterprise customers in government and defense sectors.

Capital One released VulnHunter, an open-source AI security tool that scans source code for vulnerabilities, maps exploit paths, and proposes fixes before code reaches production. Built on Anthropic's Claude Opus 4.8 model, the tool uses an 'attacker-first forward analysis' approach combined with a falsification engine to reduce false positives. Capital One's CISO Chris Nims cited the need to distribute defensive AI capabilities widely as software supply chains grow more interconnected and AI threats accelerate.

Moonshot AI, a Beijing-based startup backed by Alibaba, released Kimi K3, a 2.8-trillion-parameter open-source model that benchmarks show performs competitively with top proprietary systems from Anthropic and OpenAI. The release, timed ahead of the 2026 World AI Conference in Shanghai, represents a significant escalation in the global AI race and marks a comeback for Moonshot after losing market position to DeepSeek over the past 18 months. Full model weights are scheduled for release on July 27, with the model already accessible via kimi.com.

Canva launched Code 2.0, an AI website builder available to all 265 million monthly users including free accounts. The tool generates interactive websites from plain-language prompts and lets users edit results through drag-and-drop interfaces rather than re-prompting or modifying code. Canva is positioning itself against competitors like Lovable and Bolt.new by emphasizing design quality and editability over raw code generation.

1Password launched AI Spend and Consumption Management, a new feature within its SaaS Manager platform that tracks token consumption and spending across AI vendors including Anthropic, Cursor, and OpenAI. The tool provides IT and finance teams real-time visibility into AI costs through a unified dashboard with spend limits, alerts, and usage breakdowns by team, user, vendor, and model. The move positions 1Password in a nascent enterprise budget category as organizations struggle to forecast and control consumption-based AI costs.

Slack has integrated Slackbot with Salesforce's full platform, allowing users to query CRM data, generate Tableau charts, update records, and trigger DocuSign approvals through chat. The integration uses Salesforce's Model Context Protocol servers and Headless 360 infrastructure to create a unified conversational interface. Slack positions this as a response to competitive pressure from Microsoft Teams and Google Workspace, framing the move as part of a broader shift toward 'multiplayer AI' that keeps insights and actions visible to entire teams rather than siloed in individual agent interfaces.

Anthropic published research showing that Claude language models have spontaneously developed an internal structure called J-space that mirrors global workspace theory, a leading neuroscience model of human consciousness. Using a new mathematical technique called the Jacobian lens, researchers identified a privileged zone of internal activity where Claude holds concepts it can report on and reason with, surrounded by automatic processing it cannot access. The finding has already begun influencing how Anthropic monitors its AI systems for safety risks.

Z.ai, a Beijing-based AI lab, launched ZCode, a free desktop application designed as an agent-first development environment for its GLM-5.2 model. The tool competes directly with Cursor, Claude Code, GitHub Copilot, and Google's Antigravity in the AI coding market. ZCode's pricing undercuts competitors significantly, with plans starting at $16.20 per month, and includes features like remote control via WeChat and Feishu, reflecting the company's focus on the Chinese developer market.

Mistral AI released OCR 4, a document intelligence model that returns structured representations of documents with bounding boxes, block-type classification, and confidence scores rather than raw text extraction. The model supports 170 languages, multiple file formats, and can be deployed on-premises, positioning Mistral's European sovereignty pitch directly at regulated enterprises. OCR 4 is available through multiple platforms including the Mistral API, Amazon SageMaker, and Microsoft Foundry, with pricing starting at $4 per 1,000 pages.

Alibaba Cloud released HappyHorse 1.1, an upgraded AI video generation model now ranked No. 2 globally on independent benchmarks. The release capitalizes on market consolidation following OpenAI's discontinuation of Sora and ByteDance's indefinite shelving of Seedance 2.0 due to financial and copyright pressures. HappyHorse is positioned as an enterprise-grade, API-first product backed by Alibaba's infrastructure, targeting integration into corporate content production workflows.

Anthropic released a major overhaul of Claude Design on Wednesday that addresses the tool's token consumption problem while repositioning it as an enterprise brand-compliance layer. The update introduces design system imports that allow organizations to lock down approved components, auto-correct outputs against brand standards, and prevent individual users from overriding company guidelines. The move reflects Anthropic's broader strategy to embed Claude into enterprise workflows rather than position it as a standalone assistant.

Microsoft AI chief Mustafa Suleyman disclosed that a contractual change with OpenAI six months ago freed the company to independently pursue superintelligence using its own researchers, data, and custom silicon. The company announced seven new in-house AI models branded under the MAI family, including a flagship reasoning model and tools for coding, image generation, transcription, and voice synthesis. This marks a strategic shift where Microsoft is building alongside OpenAI rather than relying solely on it, though the company is not abandoning the partnership.

Microsoft introduced Microsoft Execution Containers (MXC) at its Build conference, an OS-level sandbox built into Windows that lets developers and IT administrators control what AI agents can access at runtime. The system enforces boundaries through the OS kernel, separates agent execution from user desktops and devices, and binds every agent to a strong identity for auditing. OpenAI and Nvidia are already on board. The move addresses a critical security gap that has hindered enterprise deployment of autonomous AI agents.

Microsoft unveiled the Surface RTX Spark Dev Box, a desktop computer featuring Nvidia's Blackwell-architecture RTX Spark processor and 128GB of unified memory, designed to run AI models with 120+ billion parameters locally without cloud API calls. The device delivers one petaflop of AI compute and will be available later this year through Microsoft.com at undisclosed pricing. The move signals a strategic shift for Microsoft, acknowledging that cloud GPU costs have become unsustainable for many development teams while betting that local prototyping will still drive Azure deployment at scale.

Perplexity AI demonstrated a hybrid local-cloud inference system at Computex 2026 that automatically routes AI workloads between a user's device and cloud models in real time, without requiring advance configuration. The system keeps sensitive data on-device while sending complex reasoning tasks to frontier models in the cloud. The feature will launch in the coming weeks on Perplexity's Personal Computer product, which runs on Intel Core Ultra Series 3 processors.

Mistral AI announced a major expansion into industrial manufacturing, a new inference data center near Paris, and a rebranded consumer assistant at its inaugural conference. The French startup, now employing 1,000 people and targeting 1 billion euros in 2026 revenue, is positioning itself as an alternative to American AI providers for enterprises concerned about data sovereignty. The company acquired Emmi AI to build physics simulation capabilities and announced partnerships with Airbus, BMW, and ASML to deploy AI across aerospace, automotive, and semiconductor design workflows.

A DataGrail analysis of 2,400 business software vendors found that 63.6% of those advertising AI capabilities do not disclose third-party AI subprocessors in their legal documentation. This means companies purchasing AI-enabled software may unknowingly expose customer data to undisclosed AI models and systems they never reviewed or approved. The finding highlights a critical gap between what vendors claim in contracts and what they actually deploy in production.

Datacurve released DeepSWE, a new coding benchmark that reveals significant performance gaps between frontier AI models previously thought to be roughly equivalent on existing leaderboards. GPT-5.5 scores 70% on DeepSWE, 16 points ahead of competitors, while an audit found SWE-Bench Pro, the industry's most cited coding benchmark, has a 32% error rate in its automated verifiers. The findings suggest enterprise procurement decisions and venture capital investments may be based on unreliable benchmark data.

Kore.ai launched Artemis, a platform that uses AI to automate the design, building, and optimization of enterprise AI agents. The system introduces Agent Blueprint Language (ABL), a YAML-based standard for defining agents, and Arch, an AI system that translates business requirements into production-ready agent systems. The move positions Kore.ai against Microsoft, Salesforce, Google, and ServiceNow in the race to become the default infrastructure for enterprise AI agents.

Resolve AI, a production-operations startup backed by Greylock and Lightspeed, announced a platform expansion featuring multi-agent investigation systems designed to diagnose production failures faster and more accurately than single-agent approaches. The company claims a twofold improvement in root cause accuracy on internal benchmarks and positions itself to address the operational debugging gap created by the AI coding boom. The announcement reflects growing tension in software development, where AI-powered code generation has accelerated shipping but left production monitoring and incident response largely manual.

Cerebras announced it is running Kimi K2.6, a trillion-parameter open-weight model from Chinese AI startup Moonshot AI, at nearly 1,000 tokens per second in production, a speed independently verified as 6.7 times faster than the next-fastest GPU cloud provider. The milestone comes less than a week after Cerebras completed a $5.55 billion IPO and directly addresses long-standing skepticism that the company's wafer-scale chips could only handle smaller models. The announcement signals Cerebras intends to compete at both the speed and scale frontier of AI inference, with enterprise customers increasingly seeking alternatives to expensive, capacity-constrained APIs from Anthropic and OpenAI.

Google has redesigned its search box for the first time in 25 years, transforming it from a simple keyword input field into a multimodal AI interface that accepts text, images, PDFs, videos, and Chrome tabs. The company is simultaneously merging its AI Overviews and AI Mode features into a single unified search experience, eliminating the friction of choosing between traditional results and AI-forward search. The redesigned box dynamically expands for longer conversational queries and includes an AI-powered suggestion system that coaches users toward more detailed questions. This shift signals Google's view of search's future as an open-ended conversational interface rather than a keyword lookup tool.

Fin, the rebranded customer service AI platform formerly known as Intercom, has launched Fin Operator, an AI agent designed to manage and optimize Fin itself. Rather than replacing human agents, Operator targets support operations teams who configure, monitor, and improve the customer-facing AI system by automating data analysis, knowledge base management, and agent debugging. The move reflects Fin's growing dominance within the company, now accounting for roughly a quarter of total revenue and virtually all growth, with the platform resolving over two million customer issues weekly across 8,000 customers.

Anthropic has surpassed OpenAI in business AI adoption for the first time, reaching 34.4% of American businesses versus OpenAI's 32.3%, according to Ramp's May 2026 AI Index. The shift represents a dramatic reversal from April 2025, when OpenAI held a commanding 32% lead while Anthropic stood at under 8%. Anthropic's surge was driven primarily by Claude Code, its agentic coding tool, which has become the company's fastest-growing product. However, the same report warns that Anthropic's lead faces three significant threats: escalating costs, compute constraints, and vulnerabilities in its token-based pricing model.

A new site called AI IQ has launched a framework for scoring frontier language models on a single intelligence quotient, mapping over 50 models onto a bell curve similar to human IQ tests. The visualization has drawn quick praise from enterprise technologists for making the fragmented model landscape legible, but sharp criticism from researchers who argue that reducing jagged, uneven AI capabilities to a single number creates false precision. The methodology groups 12 benchmarks across four reasoning dimensions (abstract, mathematical, programmatic, and academic) and applies hand-calibrated difficulty curves to prevent easier benchmarks from inflating scores, with GPT-5.5 currently leading at an estimated IQ of 136.

Anthropic unveiled three updates to its Claude Managed Agents platform at its developer conference, with the most significant being 'dreaming,' a system that lets AI agents learn from past sessions by extracting patterns and consolidating insights into auditable playbooks without modifying underlying model weights. The company also moved outcomes and multi-agent orchestration from experimental status into public beta. Early adopters report substantial gains: Harvey achieved 6x higher task completion rates, Wisedocs cut document review time by 50%, and Netflix scaled to processing hundreds of simultaneous build logs.

Miami-based startup Subquadratic emerged from stealth claiming its SubQ 1M-Preview model achieves a 1,000x efficiency gain by implementing fully subquadratic attention, where compute scales linearly rather than quadratically with context length. The company raised $29 million in seed funding and launched three products in private beta, but the AI research community has responded with skepticism, demanding independent validation of the extraordinary performance claims.

Microsoft has moved Agent 365, its AI agent management platform, from preview to general availability, signaling that enterprise governance of autonomous AI is now an operational priority rather than a future concern. The platform provides IT and security teams with centralized visibility and control over AI agents running across Microsoft's ecosystem, third-party cloud services, employee endpoints, and SaaS integrations. The launch underscores an emerging security crisis: employees are deploying AI agents without IT oversight (shadow AI), while enterprises struggle to govern agents that can access sensitive data, invoke tools, and take autonomous actions.

OpenAI converted its sold-out GPT-5.5 launch party into a monthlong developer incentive program, granting a tenfold increase in Codex rate limits to all 8,000+ developers who applied for invitations, whether accepted or rejected. The boost runs through June 5 and applies to personal ChatGPT accounts immediately. The move appears designed to convert trial usage into paid subscriptions once the promotional period ends, while the actual invite-only event proceeds as a low-key San Francisco gathering where GPT-5.5 itself reportedly helped plan the logistics.

Writer, an enterprise AI agent platform backed by Salesforce Ventures and Adobe Ventures, launched event-based triggers that enable AI agents to autonomously detect business signals across Gmail, Gong, Google Calendar, Google Drive, SharePoint, and Slack, then execute multi-step workflows without human initiation. The release includes new connectors, enhanced governance controls, and encryption key management. The move represents a shift from reactive to proactive AI agents in enterprise software, arriving as AWS, Salesforce, and Microsoft race to establish competing agentic platforms.

Netomi, a San Francisco AI startup focused on enterprise customer service automation, raised $110 million led by Accenture Ventures with participation from Adobe Ventures and others. The round is notable not for its size but for its strategic structure: Accenture is committing to a global alliance with hundreds of trained team members to distribute Netomi to its Fortune 100 client base, while Adobe plans to integrate Netomi into its Brand Concierge ecosystem. The deal signals a shift in enterprise AI from chatbot layers to deeply embedded intelligence governing entire digital experiences.

AWS launched a major suite of AI capabilities on Tuesday, including OpenAI's GPT-5.4 and GPT-5.5 models on Amazon Bedrock, a new agentic developer framework, a desktop productivity tool called Amazon Quick, and four new Connect-based AI solutions for supply chains, hiring, healthcare, and customer experience. The timing capitalized on Monday's restructuring of the Microsoft-OpenAI partnership, which replaced Microsoft's exclusive API rights with a nonexclusive license, freeing OpenAI to distribute models across rival cloud providers for the first time. AWS is positioning itself as the infrastructure layer for the agentic AI era, where intelligent agents take autonomous action inside enterprise workflows rather than just answering questions.

Mistral AI has launched Workflows in public preview, a production-grade orchestration engine built on Temporal that separates execution from control to keep enterprise data private while managing complex multi-step AI processes. The platform, which is already handling millions of daily executions, addresses what Mistral sees as the real bottleneck in enterprise AI adoption: not the models themselves, but the infrastructure to run them reliably at scale. The release comes as the agentic AI market is projected to grow from $10.9 billion in 2026 to $199 billion by 2034, yet over 40% of agentic AI projects are expected to be abandoned by 2027 due to cost, complexity, and unclear ROI.

Google launched Deep Research and Deep Research Max, two autonomous research agents built on Gemini 3.1 Pro that can now search the open web while accessing private enterprise data through a single API call. The agents support Model Context Protocol (MCP) for connecting to third-party data sources and can generate native charts and infographics in research reports. Deep Research prioritizes speed for interactive use cases, while Deep Research Max uses extended test-time compute for exhaustive background analysis. Both are available today in public preview through paid Gemini API tiers.

Cirrascale Cloud Services announced a partnership with Google Cloud to deliver Gemini as a fully private, air-gapped appliance that runs on-premises without internet connectivity. The offering packages Gemini into a Dell-manufactured hardware appliance with eight Nvidia GPUs, allowing enterprises and government agencies to access frontier-class AI without exposing data to cloud infrastructure. The product enters preview immediately with general availability expected in June or July, addressing a critical pain point for regulated industries that have struggled to balance model capability with data security.

A survey of 200 senior DevOps and SRE leaders at large enterprises finds that 43% of AI-generated code changes require manual debugging in production even after passing QA and staging tests, with zero respondents reporting high confidence that AI code will behave correctly once deployed. The findings arrive as Microsoft and Google report that roughly 25% of their code is now AI-generated, yet validation infrastructure has not scaled to match AI's production velocity. Recent high-profile outages at Amazon traced to unvetted AI-assisted code changes underscore the real-world costs of this gap.

Microsoft launched MAI-Image-2-Efficient, a lower-cost and faster variant of its flagship text-to-image model, priced 41% below the original MAI-Image-2 while delivering 22% faster inference and 4x greater GPU throughput efficiency. The model is immediately available in Microsoft Foundry and MAI Playground and is rolling out to Copilot and Bing. The rapid release, less than a month after MAI-Image-2's debut, signals Microsoft's push to build an independent AI stack and suggests its in-house superintelligence team is operating with startup-like velocity rather than traditional research lab cadence.