Cerebras is turning its wafer-scale processors into a reusable datacenter platform with the CS-4 Nexus rack, while setting a 2027 target of up to 10,000 output tokens per second per user for CS-5. The design could let operators upgrade compute without replacing the rack’s power, cooling, and I/O infrastructure, although the most ambitious performance figures remain projections.

Source

OpenAI has made ChatGPT Business Premium seats generally available, giving organizations a way to remove the Standard plan’s five-hour usage ceiling for selected employees instead of upgrading an entire workspace. Premium costs $100 per user per month with annual billing or $125 month to month, while Standard remains $20 or $25 respectively.

Source

Perplexity has launched Portable Computer, a local-first AI agent that runs its core runtime and language models on Nvidia hardware instead of sending most work to the cloud. The approach gives Perplexity Pro and Max subscribers zero token costs for tasks completed locally, while preserving the option to request cloud assistance when local models cannot finish a job.

Source

Bill Gates says governments are unprepared for an AI transition that could eliminate large numbers of entry- and mid-level jobs within a decade. He is calling for new national and international institutions, protected human-only roles, and taxes on AI and robots before economic disruption erodes public trust.

Source

Microsoft has made Intelligent Discovery in Microsoft Marketplace generally available worldwide, giving customers a conversational way to find, compare, and evaluate cloud, AI, and agent solutions. Instead of starting with a product name or keyword, buyers can describe a business problem in natural language; Microsoft says preview users were 68% more likely to find a suitable solution and proceed toward purchase.

Source

ChatGPT Work can now cross website login screens on the web and mobile, allowing its cloud browser to complete tasks inside authenticated accounts. Users enter credentials and two-factor codes directly into a secure form, while ChatGPT remains unable to see or store them—a change that expands the security boundary from public web browsing to active account sessions.

Source

ChatGPT Scheduled Tasks can now respond to activity in connected services instead of waiting for a fixed time. ChatGPT Work users can configure automations for new Gmail messages, Slack posts, and supported GitHub events, turning the feature from a reminder tool into a basic event-driven assistant.

Source

Microsoft’s August 2026 Excel update makes Copilot more useful for investigating and reshaping workbooks, adding tools for change history, chart and PivotTable creation, and Python-based analysis. The release is available across Windows, Mac, and the web, but most of the new functionality requires a Copilot license, with Python support initially limited to Excel Insiders.

Source

Microsoft is repositioning AI agents from experimental projects to managed investments, with Microsoft Foundry adding tools to track token costs, optimize workflows, and control runaway usage. The shift comes as 71% of surveyed business leaders plan to increase AI budgets, while organizations increasingly demand measurable returns rather than successful pilots alone.

Source

Claude’s memory now works across regular chats and Claude Cowork, so context built in one surface is immediately available in the other. Anthropic is also making memory update during conversations, while giving users file-level controls and an opt-in safeguard for sensitive topics.

Source

OpenAI is adding an Admin plugin that lets authorized users inspect workspace activity and change settings directly from ChatGPT Work and Codex. The conversational workflow covers membership, permissions, usage limits, and spending approvals while preserving existing roles, approval policies, and review steps.

Source

AI-assisted vulnerability analysis and exploit development are reducing the time between disclosure and attack from weeks to hours, leaving enterprises exposed while patches move through testing and change control. Microsoft argues that network-enforced, context-aware protections should provide an immediate buffer around vulnerable workloads instead of waiting for every system to be patched.

Source

Anthropic’s flagship Claude Fable 5 has captured only about 11% of corporate spending on the company’s models two months after launch, as businesses favor cheaper systems for routine coding, analysis, and automation. The weak premium uptake contrasts with Anthropic’s broader enterprise gains, suggesting that companies want access to frontier AI but are rationing its most expensive model.

Source

OpenAI is reinstating the five-hour usage limit for Codex and ChatGPT Work on ChatGPT Plus subscriptions from August 25, after temporarily relying only on a weekly cap. The change is intended to control computing demand and prevent users from consuming an entire week’s allowance during one extended session.

Source

Apple’s planned camera-equipped AirPods appear to be designed for still-image analysis rather than conventional photography: hidden macOS 26.7 code describes synchronized RGB cameras, two capture modes, and no video recording. The code also points to a status light that should alert people whenever the earbuds capture an image.

Source

Google has launched Gemini Enterprise for Legal, a preview platform that connects Gemini to law firms’ existing document, case-management, and research systems. Its agents can review contracts, investigate case law, monitor regulations, and process privacy requests while preserving access controls and keeping customer data out of Google’s model training.

Source

Azure SRE Agent is an AI-powered operations assistant that investigates incidents, analyzes telemetry and cloud resources, and helps automate routine site reliability engineering tasks. Azure is offering new customers a 30-day trial of Azure SRE Agent without the usual always-on charges, allowing teams to configure up to three agents and pay only for active Azure Agent Unit usage during the trial. Microsoft is also making VNet integration generally available and releasing Live Reports in public preview.

Source

Thibault “Tibo” Sottiaux, head of Codex at OpenAI, told Matthew Berman that ultrafast mode, voice interaction, and cloud-based agents could replace today’s cumbersome multi-agent workflows with AI that operates in real time and adapts to each user. He also described how OpenAI uses its own models to improve infrastructure, reduce costs, and automate tasks such as performance tuning and vulnerability remediation.

Source

Google is adding Gemini-powered Quick Assessments to Migration Center, allowing organizations to turn basic infrastructure data or VMware RVTools exports into an initial Compute Engine cost model in minutes. The workflow also generates service mappings, TCO estimates, and an exportable business case, potentially removing weeks of work from early cloud migration planning.

Source