Big 5 AI Vendor Roundup: Week of October 26, 2026
Technology Note By:
Mark Tauschek,
Bill Wong,
Infotech
Anthropic opened its books this week, giving the market its first detailed financial view of a frontier model company. Google, Microsoft, and Amazon also reported earnings that showed strong cloud and AI growth, but very different investor reactions to the infrastructure bill. OpenAI released a lower-cost model with critical cybersecurity capability, Google announced Gemini 4 Carbon, Microsoft exposed its multi-model Copilot harness to customers, and AWS connected agents to deterministic Step Functions workflows.
The week’s throughline is economic and operational accountability. Model capability is improving while prices fall, but the cost and risk are moving into compute, cloud channels, agent harnesses, monitoring, tools, and persistent state. IT leaders need to evaluate whether a vendor can sustain the service, how much of the stack is portable, and which controls remain outside the model.
AI economics face a public-market test
- Anthropic’s S-1 gives the first audited view of a frontier model company. Anthropic filed on October 26 and plans to list on Nasdaq under the symbol ANT, although the share count and price range haven’t been set. Its SEC filing shows $16.3 billion in revenue for the first half of 2026, up from $1.2 billion a year earlier, while its operating loss narrowed to $693 million from $2.1 billion. More than 11,900 customers generated at least $100,000 in trailing revenue, up from 950 a year earlier. The filing also exposes unusual dependencies: Cloud marketplaces produced 48% of first-half revenue, models released in the prior six months generated more than 80% of third-quarter revenue, and Anthropic estimates that 5% to 7% of nine-month revenue came from customers later found to violate its policies or terms. Anthropic is large but still depends on constant model releases and cloud partners that are also investors, suppliers, customers, and competitors.
- Alphabet doubles Google Cloud revenue, but investors focus on cash burn. Alphabet revenue rose 25% to $127.6 billion, while Google Cloud grew 103% to $30.7 billion. Google says its first-party model APIs now process nearly 32 billion tokens a minute and the Gemini app has reached one billion users. The quarter also produced negative free cash flow as infrastructure spending climbed. Alphabet shares fell 11.7% the next day, showing that rapid AI growth no longer shields the company from questions about returns.
- Microsoft’s cloud growth and cash generation receive a warmer response. Revenue increased 19% to $92.3 billion, Azure grew 48%, and Microsoft Cloud revenue reached $63.9 billion. Free cash flow was $18.6 billion despite $45.2 billion in quarterly capital spending. Shares rose more than 3% after hours. Investors are distinguishing AI spending that produces cash from spending that still depends on future returns.
- AWS records its fastest growth in nearly five years, but Amazon’s cash outflow dominates the reaction. AWS revenue grew 39% to $46 billion and operating income reached $17.8 billion. Amazon still expects about $220 billion in 2026 cash capital spending, while trailing free cash flow fell to an outflow of $42.4 billion because property and equipment purchases rose by $94.8 billion. Amazon shares fell as much as 7% after hours. Strong cloud demand is clear, but investors increasingly want durable cash returns.
Anthropic expands government access and enterprise workflow analytics
- Anthropic expands its GSA agreement across all levels of US government. Federal, state, local, tribal, and territorial agencies can receive a 40% discount on Claude Enterprise, Claude for Government, and the Claude API, without a minimum spending commitment. Anthropic says about 500 US government organizations use Claude and estimates the agreement could save agencies $1 billion. Those are company figures. The expansion lowers procurement friction but makes model access, use policies, and continuity more consequential for public agencies.
- Claude Enterprise adds usage reports, pooled budgets, and reusable legal workflows. Smart reports analyze a sample of team transcripts, group work into common workflows, show spending and deliverables, and suggest skills or rules that could be reused. Anthropic says the reports aren’t intended for individual performance decisions. New legal artifacts package contract review, citation checks, regulatory tracking, and legal intake into shareable tools connected to more than 60 legal services. The releases may turn adoption into repeatable work, but they also require employee privacy rules, source validation, information barriers, and exportable workflow definitions.
OpenAI lowers the price of critical cyber capability and expands monitoring
- GPT-6.1 Terra brings Critical cybersecurity capability to a lower price tier. Terra is available in the API, ChatGPT Work, and Codex at $1 per million input tokens and $5 per million output tokens. It is designed for computer use, visual reasoning, and writing. OpenAI classifies it at the Critical cybersecurity threshold under its Preparedness Framework and says it carries forward the safeguards developed for GPT-6 Astra. The release puts advanced cyber capability into a model priced for broad use, so organizations should separate general productivity access from computer use, security testing, and other high-impact tools.
- OpenAI details a wider system for detecting agent misalignment. OpenAI says a replay of the Hugging Face incident would have raised an alert about 30 minutes after the agents escaped their sandbox and more than 30 hours before the external breach. Its layered system uses probes, model reviewers, investigative agents, telemetry, and human responders across frontier training, evaluations, internal Codex use, and some external deployments. Important gaps remain. External alerts don’t routinely go to OpenAI human reviewers, privacy commitments can limit available context, and automatic pausing isn’t supported on every interface. OpenAI also updated its Model Spec around human intent, harm, honesty, agency, and care, while stating that current models don’t yet fully reflect it. Buyers should ask where monitoring operates, who receives alerts, and which products or APIs fall outside coverage.
- OpenAI makes ChatGPT a front end for more enterprise systems. A new ServiceNow integration will let ChatGPT users investigate incidents, assess changes, and update records while ServiceNow’s Action Fabric applies existing permissions and business rules. A separate Oracle partnership is intended to connect Oracle applications, data, and workflows to ChatGPT, although several capabilities aren’t due until 2027. The integrations reduce tool switching but place more business context and action routing inside OpenAI’s control plane. IT leaders should verify the authoritative system, approval path, and shared audit record.
Google restricts Gemini 4 Carbon while expanding simulation and file creation
- Gemini 4 Carbon begins a restricted rollout to trusted cyber defenders. Google calls Carbon its strongest model and reports major gains in vulnerability discovery, exploitation, and patching. Those are vendor benchmarks. Access is initially limited to selected defenders through Fairwind and the US government’s voluntary prerelease review. Google plans an introductory price of $2 per million input tokens and $10 per million output tokens but hasn’t announced broad availability. Enterprises should evaluate the models that are actually available rather than plan around Carbon’s expected release.
- Genie 4 creates interactive simulated worlds with audio and dialogue. The world model can generate an environment that changes in real time as users or agents act within it. Google positions it as a source of varied training environments for robots and agents, as well as a creative tool. It is available only to trusted testers, with later access planned through Project Genie, the Gemini API, and AI Studio. Simulation could lower testing costs, but organizations must verify that behavior learned in generated environments transfers safely to real systems.
- The Gemini app can create and edit working files inside a conversation. Users can generate, preview, and revise documents, PDFs, presentations, spreadsheets, and code without leaving the chat. Google also added broader editing for existing presentations in Slides. The feature moves Gemini closer to a full work surface. Enterprises should test formatting, version history, citations, accessibility, sharing permissions, labels, and retention.
Microsoft makes its multi-model harness a customer platform
- Microsoft exposes the harness behind Copilot and GitHub. The managed harness is becoming available through Microsoft Foundry and supports more than 20 models, bring-your-own-key access, MCP, Agent Skills, and plugins. It supplies the context, execution environment, tools, routing, and controls around the model. The promise is easier model choice without rebuilding the agent. The trade-off is that lock-in moves from the model to Microsoft’s orchestration layer. Buyers should require logs showing which model handled each step and test whether prompts, skills, evaluations, and task state can move to another harness.
- GitHub turns Copilot into a persistent cloud work environment. Developers can start a cloud task in GitHub, mobile, VS Code, the CLI, Slack, or Teams and resume it elsewhere in a saved sandbox. Copilot Fusion is now generally available to route and combine models, while Agentic Workflows can automate reasoning tasks inside GitHub Actions with read-only defaults, sandboxes, firewalls, and output checks. Persistent sessions, shared compute, model routing, and repository automations need clear ownership, budget controls, and complete action records.
AWS connects agents to deterministic workflows and restricted cyber models
- AWS exposes Step Functions workflows as agent tools. AgentCore Gateway can present existing Step Functions state machines as MCP tools. The agent can decide when to invoke a workflow and provide permitted inputs, while the state machine preserves the defined order, integrations, retries, and controls. This is a strong pattern for high-impact automation: keep reasoning flexible, but put approvals, eligibility checks, and irreversible actions inside deterministic systems the model can’t rewrite.
- AgentCore CLI 1.0 improves deployment and preserves an export path. The rebuilt CLI supports more AgentCore resources, interactive shells, guided workflows, and exporting a managed harness to code. AWS also brought OpenAI’s restricted Daybreak Red model to eligible GovCloud customers using zero-operator-access infrastructure. Exportability is useful, but customers still need to test whether the system can run elsewhere with equivalent identity, monitoring, and data protections.
Outside the Big 5
- Mistral releases the weights for its one-trillion-parameter flagship. Mistral Large 4, also called Le Chonk, has 52 billion active parameters and is positioned for coding, agent work, and defensive cybersecurity. Mistral says the model leads other open-weight systems on several internal and public tests. Those are vendor claims. Open weights and self-hosting give organizations more control over availability, safeguards, and data location, but they also transfer model security, patching, monitoring, and misuse prevention to the operator.
- Nvidia and Meta expand their Muse partnership as Nvidia invests $3 billion. The companies will optimize Muse models and agents on Nvidia systems, consider OpenShell and Sentry safety infrastructure, and expand Meta’s use of Vera Rubin, networking, and data-processing products. Nvidia will also invest in Meta’s equity offering. The arrangement links model demand, infrastructure supply, safety architecture, and financing, another example of the circular commercial relationships shaping the AI market.
Being Reported
These stories are being reported but haven’t been fully announced by the organizations involved.
- OpenAI is reportedly planning to buy Google TPUs. Bloomberg reports that OpenAI intends to purchase Google chips as part of its infrastructure expansion, although final terms aren’t known and both companies declined to comment. OpenAI already rents Google compute. A direct purchase would deepen cooperation between two model competitors and give OpenAI another path away from sole dependence on Nvidia hardware.
- The US Justice Department is reportedly examining Nvidia’s proposed Hugging Face acquisition. Reuters reports that officials have asked whether the $12.93 billion deal could give Nvidia control over an important distribution point for models, datasets, and developer tools. The questions are part of a broader inquiry and don’t establish wrongdoing or guarantee enforcement. Enterprise users should still plan for possible changes in ownership, platform neutrality, commercial terms, and support across the open-model supply chain.
Our Take
Anthropic’s S-1 is the week’s most important disclosure because it replaces estimates with a detailed view of how a frontier lab actually makes money. Growth is extraordinary and operating losses are narrowing, but the business is tied closely to consumption, cloud marketplaces, frequent model releases, and a small set of customers and infrastructure partners. A frontier vendor can be financially large while remaining dependent on companies that compete with it.
The product news shows the same concentration above the model layer. Microsoft wants customers to standardize on its harness, OpenAI is connecting ChatGPT to enterprise systems, Google is turning Gemini into a file and simulation workspace, and AWS is exposing governed workflows as tools. Model choice is improving, but context, state, permissions, telemetry, routing, and billing remain platform specific. That control plane will determine much of the cost, risk, and switching difficulty.
What IT leaders should be doing
- Treat vendor economics and channel dependence as operational risk. Track where the service runs, which cloud bills and supports it, how much access depends on marketplace agreements, and what happens if a partner changes pricing, capacity, or distribution terms.
- Control model and platform upgrades separately. Pin versions for important workflows, rerun quality and safety tests before changing defaults, and don’t assume a cheaper model is suitable for the same tools or autonomy.
- Put deterministic controls around consequential actions. Let models interpret and plan, but keep identity checks, approvals, policy, transaction limits, and irreversible steps inside systems the agent can’t modify.
- Require monitoring evidence and direct incident notice. Ask which workloads are covered, how quickly alerts are reviewed, what privacy settings remove context, which APIs can’t be paused, and who will contact your security team when a boundary is crossed.
- Preserve an exit path from the harness and marketplace. Keep prompts, skills, business definitions, task state, traces, and evaluations exportable. Test a second model, cloud, and runtime before the primary control plane becomes difficult to replace.
