Need to know
- Microsoft Copilot relaunches as a unified 'super app' merging chat, coding, and always-on Autopilot agents for business.
- OpenAI is days away from previewing GPT-6 Cyber, a cybersecurity-focused model with a novel secure-deployment product.
- Google Gemini gains 'Call for Me,' letting the assistant make autonomous phone calls on Pixel 11 for Gemini subscribers.
- SentinelOne extends Wayfinder threat hunting across AWS, Azure, and Google Cloud, unifying endpoint, identity, and cloud investigation.
- Anthropic graduates Claude Code Cloud Sessions from experimental to paid-plan GA, backed by $100–$250 in launch credits.
New Releases
Microsoft today shipped a redesigned Copilot that merges the consumer and enterprise versions into one app with three modes: Home for chat, Code for software development, and Autopilot for always-on background agents.
- Home and Code tabs reach Frontier program users first; Autopilot enters private preview by month-end, positioning Copilot as a paid M365 business offering rather than a free chatbot.
- Microsoft is rebranding Scout into the Autopilot surface, signaling that personal-agent functionality is now a core enterprise product rather than a consumer experiment.
OpenAI will preview GPT-6 Cyber, a cybersecurity-specialized model, within days, accompanied by a first-of-its-kind secure automated deployment product for enterprise customers.
- A small cohort already has early access through the application-only Daybreak Red program, suggesting OpenAI is using a controlled rollout to validate red-team-grade outputs before broad release.
- The secure deployment product would be OpenAI's first infrastructure play for hardened enterprise environments, directly competing with Anthropic's Glasswing PAM integration announced this week.
Google on September 24 launched 'Call for Me,' a Gemini feature that autonomously dials real businesses on a user's behalf to ask questions or make reservations, available now as an experiment for U.S. Pixel 11 owners with a Gemini subscription.
- Users can monitor the call live and take over at any point, which lowers the trust bar compared to fully opaque agentic actions and may accelerate adoption in enterprise scheduling workflows.
- The feature requires Pixel 11 and a paid Gemini subscription, narrowing the immediate audience but establishing a defensible hardware-software moat that neither OpenAI nor Anthropic can replicate today.
Anthropic on September 23 made Claude Code Cloud Sessions generally available for all paid plans, running coding tasks in Anthropic-managed VMs and dropping $100 and $250 promotional credits for Pro and Max subscribers respectively, valid through October 7.
- Sessions can start from claude.ai/code, the mobile app, or the desktop app, removing the local-machine dependency that limited Claude Code to developers with always-on hardware.
- The promotional credit window ends October 7, creating a short adoption funnel that will give Anthropic real usage data on cloud-hosted agentic coding before permanent pricing is set.
Funding
Cohere on September 25 opened Compass Cloud, a fully managed retrieval pipeline and inference service built on its Compass platform, to private beta — a product bet that Cohere can monetize RAG infrastructure as a service rather than selling only models, relevant as its Aleph Alpha merger nears close.
SentinelOne extended Wayfinder Threat Hunting to AWS, Azure, and Google Cloud, coupling Singularity Platform telemetry with Google Threat Intelligence in one workflow — a capability expansion that makes SentinelOne's managed hunting service directly competitive with CrowdStrike's Falcon Complete in multi-cloud SOC environments.
Replit acquired Atta, a business analysis and charting startup, to give non-technical users interactive data exploration inside Replit's chat interface — a move that pushes Replit's addressable market from developers toward the business analyst and operator tier that tools like Notion and Salesforce currently own.
Case Studies
Perplexity replaced Amazon DynamoDB with CobbleDB, an internally built Rust key-value store, cutting query latency fivefold and meaningfully reducing cloud storage costs for its LLM search serving tier.
- The migration was driven by multi-kilobyte document batches served to LLMs under heavy query volumes, a pattern DynamoDB's latency and cost model handled poorly at Perplexity's scale.
- CobbleDB is not open-sourced, meaning this is a competitive moat story rather than an ecosystem contribution — Perplexity is betting infrastructure differentiation compounds into product differentiation.
Silverfort integrated with Anthropic's Compliance API to give security teams a real-time unified view of Claude users, credentials, and activity across enterprise deployments, addressing the shadow-AI access problem where API keys and MCP connections proliferate without central visibility.
- The integration surfaces Claude identities alongside existing identity data, letting SOC teams answer 'who has access and what can they do' without stitching together admin logs and offboarding records manually.
- Anthropic's Compliance API is becoming an ecosystem control point: this is the second major security vendor this week to build on it, after Delinea's Glasswing integration announced Monday.
Trending on X
- Microsoft Copilot router model opacity Ethan Mollick and others are flagging that Microsoft's new Copilot routes requests through mystery models, warning that opaque model selection systematically underestimates task difficulty and degrades output quality in ways users cannot detect.
- Meta Muse overtaking ChatGPT in app charts The AI and tech community is reacting to Muse topping App Store charts, with Scale AI's Alexandr Wang posting screenshots of increasingly high-status use cases, framing it as the first genuine consumer threat to ChatGPT's brand dominance.
- Frontier LLM cost per agent decision Developers are debating the economics of using frontier models for trivial agent decisions, with critics pointing out that paying three cents per call for a model to choose between three options makes most agentic architectures financially absurd at scale.
- AI tool abstraction and engineering difficulty François Chollet's post arguing that software engineering difficulty stays constant regardless of abstraction level is generating pushback and agreement in equal measure, with the debate centering on whether AI coding tools expand output or just shift cognitive load.
- Anthropic vs OpenAI IPO timing Prediction market data showing Anthropic at 93% probability to IPO before OpenAI is circulating widely, with the community treating it as a meaningful signal given both companies' recent model release velocity and cost reduction moves.