Big 5 AI Vendor Roundup: Week of November 2, 2026

Frontier capability moved across two important thresholds this week. OpenAI released GPT-6 Cyber, a restricted model it classifies at its Critical cybersecurity level. Anthropic told investors that future models may soon meet its definition of dramatically accelerated AI research and development, while an independent assessment estimated that its model progress on well-specified tasks is running at about twice its earlier pace.

Commercial pressure is rising at the same time. Anthropic disclosed nearly $900 billion in compute commitments, OpenAI reportedly closed part of a smaller-than-planned funding round, Microsoft began limiting some unlicensed Copilot Chat use, and data center opposition became a more visible constraint. IT leaders should treat capability, vendor economics, infrastructure supply, and controls as one risk picture.

OpenAI releases a Critical cyber model and scales vulnerability discovery

• GPT-6 Cyber moves advanced exploit development into restricted production use. OpenAI released GPT-6 Cyber on November 5 through Daybreak Red for selected defenders. It classifies the model at the Critical cybersecurity threshold and reports a 48.5% score on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. OpenAI says the model stayed within scope in 99.93% of tests without its deployment prompt and 100% with it. These are vendor results. Advanced exploit development now requires separate access, monitoring, network, and approval controls from general productivity AI.
• OpenAI turns vulnerability research into an industrial pipeline. OpenAI says its autonomous Linux campaign reproduced 1,724 vulnerabilities, created working privilege escalation exploits for 948, and sent 304 critical findings to maintainers with candidate patches. The model also helped find 81 previously unknown high-severity Firefox vulnerabilities. Human review and coordinated disclosure remain necessary, but the volume changes the economics of defense and attack. Enterprises need to shorten the time between a patch becoming available and reaching production.
• OpenAI plans to acquire Pierre to strengthen Codex. Pierre Computer Company built developer tools for code storage, repository trees, diffs, and code highlighting. OpenAI plans to integrate that technology into Codex while keeping Pierre’s open-source projects available. The acquisition adds more of the development environment to OpenAI’s stack. Buyers should test whether code state, review history, and workflow metadata remain portable.
• Images 2.6 adds faster generation and more developer controls. OpenAI says the model is up to 50% faster than Images 2.5 and adds GIF creation, image markup, SVG export, more precise editing, and new API controls for complexity and rendering. It is available across ChatGPT, Codex, and the API. Enterprises still need brand, rights, provenance, accessibility, and approval controls before generated media reaches customers.
• The Grid Protectors Program ties future compute campuses to utility demand response. OpenAI says future campuses will pay their own infrastructure and electricity costs and reduce consumption when local grids are under strain. Its Georgia agreement includes up to 1,000 megawatts of flexible demand response. The program responds to growing concern over data center effects on electricity prices and reliability. Capacity plans now depend on local policy, grid agreements, and community acceptance as well as chips and capital.

Anthropic says AI may be accelerating its own research

• Anthropic’s amended S-1 raises its compute commitments and its safety warning. The November 2 filing reports $17 billion in third-quarter revenue and $895.4 billion in compute and infrastructure commitments, including $69.8 billion due over the next 12 months. Anthropic says future models may soon meet its Responsible Scaling Policy threshold for dramatically accelerated progress attributable to AI. It delayed pretraining a frontier model for more than a month to strengthen model weight security and is considering mandatory human approval for some research decisions.
• An independent assessment estimates that Anthropic’s model progress has roughly doubled. METR estimates that Anthropic is improving performance on well-specified tasks at about twice the rate seen from March 2024 through October 2025, with a plausible range of 1.3 to three times. It judges that most of the current acceleration is likely attributable to AI use in research and development. Anthropic gave METR broad access but redacted most of the evidence METR considered most informative. The result is a warning signal, not a precise measure of recursive self-improvement.
• FlashProve offers cryptographic evidence that a response came from the claimed model. Anthropic’s open-source prototype lets a provider commit to a model and generate a proof that each response came from it, rather than from a cheaper substitute or a tampered output stream. The current system carries roughly 70 to 95 times normal inference overhead, so it isn’t ready for routine use. The direction is important as buyers may eventually require technical proof of model identity rather than routing logs and contract terms.
• Claude Enterprise adds security posture reviews and pre-action policy checks. A new security center compares settings with Anthropic’s recommended configuration, while generally available inference hooks let data loss prevention and security tools inspect prompts and proposed tool calls before they run. Anthropic also added centralized administration, customer-managed IP allowlists, session termination through an API, controls that block personal Claude accounts on managed desktops, and endpoint detection inside the local Claude virtual machine. These controls move enforcement closer to the enterprise, but the security center is guidance, not certification.
• Claude adds read-only access to personal financial accounts. Anthropic began rolling out Personal Finance in Claude to US subscribers on Pro, Max, and Ultra plans. Users can connect bank and credit card accounts through Plaid, then ask Claude to categorize spending, identify upcoming bills, explain interest charges, and flag duplicate payments, unexpected fees, or rising expenses. Claude can access balances, credit limits, and up to two years of transactions for selected accounts, but it can’t move, manage, or invest money, and it doesn’t receive the user’s bank password. Anthropic says conversations in the Personal Finance hub aren’t used to train its models. The beta isn’t available on Team or Enterprise plans, but it shows how quickly assistants are moving into highly sensitive financial data, where consent, accuracy, retention, and account disconnection controls will be critical.
• Claude Marketplace grows into a channel for partner agents and products. Anthropic added security, developer, productivity, and creative offerings from more than 25 partners. Eligible customers can apply part of their committed Anthropic spending to these products. That simplifies procurement but turns one vendor commitment into a broader ecosystem budget. IT leaders should review publishers, permissions, data paths, support, and exit terms.

Microsoft shows how AI compresses the vulnerability window

• Microsoft observes agents harvesting exposed secrets at scale. Microsoft says threat actors used an agentic offensive framework to submit large volumes of GitHub queries for credentials beginning in August. In at least one follow-on campaign, stolen application credentials were used to access Entra ID tenants and connected Azure and Microsoft 365 resources. Publicly exposed secrets must be treated as already discovered. Organizations need automated scanning, rapid revocation, workload identity controls, and monitoring for unusual application logins.
• Windows vulnerability discovery accelerates, but enterprise patching remains slow. Microsoft’s October security update addressed more than 1,100 Windows CVEs as AI assisted discovery, remediation, and testing increased the volume of findings. Microsoft says fewer than half of devices in a measured enterprise sample had installed the August update after two weeks, and about one in four remained unpatched after four weeks. As models reduce exploit development from days to hours, month-long patch cycles are no longer defensible.
• Microsoft expands discovery and containment for local AI agents. Defender can now discover more than 50 supported local agents and their configured MCP connections across Windows and macOS. Microsoft Execution Containers provide operating system controls on files, networks, processes, and user interaction. Inventory is only the first step. Each agent still needs an owner, constrained identity, and telemetry connecting the agent, user, device, tools, and actions.
• Microsoft adds dynamic limits to Copilot Chat for users without a paid Copilot license. The limits began November 2 for organizations with 2,000 or fewer Microsoft 365 seats and start November 9 for larger organizations. Microsoft doesn’t publish the thresholds, and administrators can’t view or adjust them. Users may be prompted to request or buy a Copilot license when they reach a limit. IT leaders should expect help desk demand and license pressure, and base upgrades on measured workflow value rather than limit notifications.
• GitHub puts coding agents into more collaboration and development surfaces. @GitHub in Slack and Teams is now generally available for investigating bugs, creating issues, changing code, and opening pull requests. GitHub also added Copilot to Xcode and JetBrains Air, expanded agent usage metrics, and enabled Dependabot updates for packaged agent skills, prompts, and instructions. The wider reach adds more places where repository permissions, tool approvals, extensions, and agent-generated changes must be governed.
• GitHub Copilot carries coding sessions across devices and keeps working in the cloud. The generally available experience lets developers start a task in GitHub Mobile, GitHub.com, VS Code, the Copilot app, the CLI, Slack, or Teams, then continue the same session from another device. Work runs in an isolated GitHub cloud sandbox, so it can continue after the developer closes a laptop, and saved snapshots let the session resume later with its context and files intact. The flexibility is useful, but it also creates persistent cloud development environments that need repository restrictions, single sign-on, tool policies, activity logs, and compute spending controls.
• Microsoft Foundry adds managed multistep agent execution and recovery. The Microsoft Copilot harness, now in public preview in Foundry Agent Service, lets agents read and manage files, run code, invoke approved tools, apply reusable skills, and delegate bounded work to subagents. Long running resilience is generally available for hosted agents and can restart interrupted work using the original input. Developers still need to save meaningful checkpoints and prevent external actions from being repeated after recovery. IT leaders should treat recovery as part of the workflow design, with idempotency controls, approval boundaries, retained state, and complete action logs.

Google lowers video costs and brings analysis into Forms

• Gemini Omni 1.2 Flash creates longer video at a lower price. The generally available model generates clips up to 15 seconds, supports conversational editing, improves reference consistency, and costs $0.07 per second for 720p video, 30% less than Omni 1.1. It is available through the Gemini API, Gemini for business, Google Cloud, and Google Vids. Enterprises should validate consent, rights, consistency, watermark retention, and the cost of retries and upscaling.
• Gemini in Forms writes and runs Python to explain survey responses. The feature identifies patterns, compares answers across questions, summarizes findings, and recommends next steps. Because calculations run against actual responses, it may be more useful than a generic summary. Organizations should still verify the code, sample size, assumptions, and use of sensitive data.
• Google previews policy-based client-side encryption for sensitive files. Administrators can use Workspace data protection rules to encrypt matching Docs, Sheets, Slides, Office files, PDFs, and images in place while preserving the original URL. The beta requires Assured Controls and an external key service. This strengthens the boundary for sensitive content, but encryption may limit Gemini and collaboration features. Security and productivity policies need to be tested together.

AWS turns current web research into a managed agent tool

• Web Answer returns a synthesized, cited response through one AgentCore call. The generally available connector plans searches, runs several in parallel, decides when it has enough evidence, and returns an answer with citations. It removes the need to build a separate search and synthesis pipeline but concentrates source selection inside AWS. Buyers should test coverage, freshness, citation accuracy, blocked domains, and disagreement handling.
• AgentCore Runtime V2 expands and supports more connection-heavy agents. V2 is now available in five additional regions and supports up to 250 concurrent connections per session with the default compute allocation, up from 16. AWS also added a platform latency metric to separate runtime overhead from agent processing. These controls help streaming and parallel tool use but can also increase consumption quickly. Organizations need per-agent limits, cost attribution, and load tests.
• AWS expands OpenAI API support and cross-region routing in GovCloud. On October 30, Amazon Bedrock added GPT-5.6 Terra and Luna to its standard runtime endpoint in AWS GovCloud, with support for the OpenAI Responses and Chat Completions APIs as well as Amazon’s Converse API. Cross-Region Inference can automatically route requests between the US-East and US-West GovCloud regions to increase throughput and provide more capacity during demand spikes while keeping processing inside the US GovCloud boundary. Usage is captured through Bedrock invocation logs, CloudWatch metrics, Cost Explorer, and AWS cost reports. The update gives government customers a more resilient and familiar way to deploy OpenAI models, but agencies should confirm that policies permit processing in both GovCloud regions and test how routing affects latency, cost, and application behavior.

Outside the Big 5

• Marvell agrees to acquire Lumentum for $115 billion. The all-stock deal combines Marvell’s custom chips and networking technology with Lumentum’s optical components and manufacturing. The companies expect more than $1 billion in annual savings and plan to close in the first half of 2028, subject to approval. The deal shows AI constraints shifting from accelerators into networking, optics, packaging, and manufacturing.

Being Reported

These stories are being reported but haven’t been fully announced by the organizations involved.

• OpenAI reportedly closed $11 billion of a reduced funding round. Bloomberg reports that the initial close values OpenAI at $1.25 trillion and that the round could reach $20 billion. OpenAI had initially sought as much as $30 billion at a higher valuation. Capital remains available, but investors are placing a higher price on infrastructure spending and political risk.
• Anthropic reportedly plans to proceed with its IPO at a lower valuation. The Financial Times reports that Anthropic still intends to list before the US Thanksgiving holiday and may raise $80 billion to $100 billion at a valuation between $1.25 trillion and $1.75 trillion. The range remains unsettled, and Anthropic hasn’t announced the share count or price.
• Isomorphic Labs is reportedly discussing a $5 billion round at a $50 billion valuation. Bloomberg says the Alphabet-backed drug discovery company is still in preclinical development and that the financing hasn’t closed. The reported valuation prices a proprietary AI research platform before clinical evidence is available.

Our Take

The GPT-6 Cyber release and Anthropic’s acceleration disclosures belong in the same conversation. OpenAI now has a model that can produce working exploits at industrial scale, while Anthropic and METR say AI is materially increasing the pace of model research. Both vendors are adding safeguards, access limits, external review, and human gates. Evidence still comes largely from the companies, and METR’s public report omits much of the evidence behind its conclusions. A model label or system card isn’t complete assurance.

The economic signals are also becoming harder to separate from operational risk. Anthropic has nearly $900 billion in compute commitments, OpenAI raised less than it initially sought, Microsoft is using dynamic usage limits to push paid Copilot licenses, and data center capacity faces greater grid and political constraints. A model may be portable while its runtime, marketplace, identity, memory, and terms remain tied to one vendor. Resilience requires another operating path, not another model in the same control plane.

What IT leaders should be doing

• Separate high-risk capability from general access. Put cyber models, computer use, external browsing, code execution, and other consequential tools behind named identities, explicit scope, isolated environments, short-lived credentials, and human approval.
• Demand evidence for release and research controls. Ask how vendors measure AI-assisted research acceleration, what evidence independent assessors can publish, what remains redacted, and which findings can delay training or deployment.
• Close the patch gap. Use automated inventory, staged deployment, hotpatching where available, and risk-based service targets. Assume that public vulnerabilities will be analyzed by AI within hours.
• Model the commercial failure modes. Track usage limits, marketplace dependencies, compute commitments, data center constraints, and funding risk. Test what happens when access is throttled, a price changes, or a vendor delays a planned service.
• Keep the agent supply chain visible. Inventory agents, MCP servers, skills, plugins, extensions, and research connectors. Approve publishers, pin versions, retain action records, and test whether state and policy can move to another runtime.

Want to Know More?

• Big 5 AI Vendor Roundup: Week of October 26, 2026
• Big 5 AI Vendor Roundup: Week of October 19, 2026
• Overlapping AI Outages Expose an Enterprise Resilience Gap