AI intelligence

Google's September AI roundup pairs new Gemini models with wider app integrations. The company says Gemini 4 Argon has a million-token context window and is rolling out first to trusted cyber defenders through Fairwind, with broader access planned after feedback on safeguards. Other announcements include Gemini on Windows, more Connected Apps, and WeatherNext 3 forecasts integrated into Search, Gemini, Maps and Cloud.

Web Search API is now available in beta. Web Search API lets your AI agents and applications search the Internet and ground their responses in live information, instead of guessing URLs or relying on a model's training cutoff. At launch, you can choose between three search providers: Ceramic.ai, Exa, and Linkup. All three support Zero Data Retention for requests made through Cloudflare, and all have committed to Cloudflare's verified bot crawling standards. Web Search API runs through AI Gateway, so search requests appear in your gateway logs and are billed to your AI Gateway credits at each provider's list API price, with no additional markup. You can also bring your own provider API key.

Vercel says financial AI company Rogo is rebuilding internal applications on its platform and deploying agent-written code in five minutes. The customer case study reports more than 73,000 deployments in one month and six production AI agents handling work including churn analysis and deal support. Vercel also describes automated production-incident triage and remediation using its AI SDK.

GitHub has opened a public preview of computer use in Copilot CLI and the Copilot app for macOS and Windows. The assistant can work through desktop interfaces, including software without an API or command line. GitHub says users approve app control and can review those permissions; managed organizations can disable the feature. On macOS, additional accessibility and screen-recording permissions are required.

GitHub’s September roundup covers Copilot changes across VS Code versions 1.136 through 1.140. Preview features add recurring agent tasks and assistance with preparing pull requests for merge, while session controls and Dev Container support aim to keep work organized. The update also expands ways to attach GitHub issues and pull requests to chats. Availability varies by feature; several additions remain previews.

GitHub introduced dynamic workflows in public preview across Copilot CLI, the Copilot app and the Copilot SDK. Workflows define agent orchestration in code, combining sequential or parallel steps, structured results and checkpoints for review. They are available on all Copilot plans; the app needs no setup, while the latest CLI requires experimental features to be enabled.

Cloudflare has introduced Clef and Clef-flash on Workers AI for applications that need structured decisions instead of generated prose. The models return probabilities for defined answers, supporting tasks such as routing requests or escalating cases to people. Cloudflare says it has released the weights under Apache 2.0 and is seeking design partners for a reinforcement-learning fine-tuning service.

Vercel added Convai Innovations’ Laya decision model to AI Gateway for tasks such as routing requests, evaluating guardrails and scoring responses. The model returns structured decisions with probabilities. Calls served by Boundless are free through October 31, 2026; after that promotion, the dedicated free model ID stops serving, while the standard ID becomes billable. Applications should account for that difference before deployment.

Version 1 of Cloudflare's Workers OAuth Provider separates the authorization server from the MCP resource server, allowing them to run in different Workers. Tokens can be checked through a Service Binding rather than over the public internet. The release supports the July 2026 MCP authorization specification while retaining compatibility with older clients, including dynamic client registration.

Vercel’s AI Gateway now supports Microsoft’s MAI-Voice-2.1 and lower-latency Flash speech models, plus MAI-Transcribe-2 Streaming for live transcription. Vercel says the speech models support 23 languages and the transcription API returns updates as audio arrives. Partial transcripts may change as more audio is processed. The announcement also states that inference is billed at listed model rates without an additional platform markup.

Vercel Agent can now install private dependencies from npm and custom registries using team-shared environment settings. The change supports npm, pnpm and classic Yarn, with separate configuration for npm tokens or additional registries. Vercel says credential values stay outside the agent’s sandbox. Project-scoped variables are not used for this feature, so existing build configuration may need to be adjusted.

Vercel now offers Browserbase Search and Fetch through AI Gateway, letting supported models locate current web pages and retrieve their contents. The same tools can be retained when switching model providers that support tool calling. The integration requires AI SDK 7.0.116 or later and uses an AI Gateway key; the announcement directs developers to separate pricing and configuration details.

Google has announced Gemini 4 Argon, with initial access for selected cybersecurity defenders through its Fairwind Program. The company says the model supports longer reasoning workflows, including software engineering and enterprise tasks, with an output limit of one million tokens. Broader availability is planned after additional testing and safeguards work; the reported capabilities and benchmark results are Google’s claims.

Google Cloud introduced a preview remote MCP server that exposes gcloud and BigQuery command-line operations to AI agents. Commands run in an isolated cloud environment instead of requiring local CLI installation. Google says execution uses the authenticated caller’s permissions and organization policies, with optional audit logging and Model Armor integration. The preview is intended to simplify infrastructure and data workflows while retaining existing access controls.

Google Cloud’s September infrastructure roundup focuses on running larger fleets of AI agents. It introduces GKE Agent Substrate, adds scale-to-zero and custom-metric autoscaling capabilities, and describes pod snapshots that preserve workload state. Storage updates include preview Filestore agent volumes and new virtual-machine options. The pod-snapshot performance claim comes from Google’s internal testing, so results may differ across workloads.

Google Cloud has made Spanner Omni generally available for deployment on customer infrastructure or other clouds. It combines relational, graph and vector capabilities with production security, backup and support options. Unlike managed Spanner, customers operate the infrastructure themselves, and Google does not provide an availability SLA. A free development edition and paid commercial edition have different usage and feature conditions.

Google DeepMind introduced SynthID Bio, a research approach for marking AI-generated protein sequences and predicted structures so their provenance can be checked. The team reports laboratory tests in which watermarked protein designs retained biological function. It presents the work as an additional biosecurity and scientific-integrity measure, while acknowledging that stronger resistance to deliberate tampering remains an open research challenge.

GitHub is extending its HydraFusion research preview to VS Code and the Copilot app. Instead of selecting only a model, the system can choose a single-model, draft-and-escalate or independent-critique workflow. Access depends on the supported Copilot plan and preview settings; GitHub says the research preview remains subject to change.

Google Cloud’s Data Agent Kit is now generally available, combining MCP tools with Google-authored skills for coding agents. The release adds support for BigQuery Graph, Bigtable and Managed Service for Apache Spark, helping agents inspect schemas, query data and work with operational context. It integrates with editors and command-line agents, while each operation runs under the user’s existing IAM permissions.

Vercel added InclusionAI’s Ling 3.1 Flash to AI Gateway for coding, analysis and tool-using agents. The announcement describes a hybrid reasoning model with a 262,000-token context window on the service. Access is promotional and free through October 13, 2026. Afterward, the standard model identifier begins billing, while the separate free identifier stops serving rather than automatically incurring charges.

AI Gateway now supports Machine Payments in beta. With Machine Payments, clients can use the x402 protocol to pay for eligible inference requests directly from a stablecoin wallet instead of maintaining a prepaid credit balance. Machine Payments is available for the /ai/run endpoint with select open models. To request x402 payment, authenticate with a Cloudflare API token and include the Cloudflare-specific Payment-Method: x402 header: curl -iX POST "https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run" \ --header "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \ --header "Payment-Method: x402" \ --header "Content-Type: application/json" \ --data '{ "model": "z-ai/glm-4.7-flash", "input": { "messages": [ { "role": "user", "content": "What is Cloudflare?"

GPT-6.1 Sol is rolling out in GitHub Copilot for supported paid plans across editor, terminal, mobile and coding-agent surfaces. GitHub says availability is gradual and enterprise administrators can control access through model policies. The announcement describes usage-based billing at provider list pricing; its performance statements are GitHub's early testing claims.

Google made its reinforcement-learning version of GKE Agent Sandbox and orchestration SDK generally available. It combines prewarmed environments, image streaming and reusable pods to reduce time that training accelerators spend waiting for execution sandboxes. In Google’s tests on a fixed ten-node pool, average startup improved roughly tenfold and worst-case waits improved up to 45-fold; those figures describe the published benchmark configuration.

Google Cloud expanded its Gemini Enterprise catalog with partner-built security agents and integrations for protecting AI workloads. Announced capabilities span identity containment, sensitive-data controls, application-security triage and defenses against prompt injection. The examples include workflows that retain human approval for privileged actions. These are vendor and partner product claims, rather than evidence that deploying the integrations eliminates security risk.