AI intelligence

Vercel reports that TypeSafe AI’s Jev reached nearly 13% of AI Gateway’s paid teams within its first day, more than twice the adoption of any previous model launch on the service. Jev is designed for software decisions, returning typed answers and probabilities instead of prose. The report describes early uptake and leaves the durability of that adoption as an open question.

Vercel added GLM 5.3 FlashX, a faster serving option for Z.ai’s multimodal coding model, to AI Gateway. Vercel describes inference at roughly 200 tokens per second, targeting interactive applications and repeated agent tool calls where response speed matters. Developers can select the model through supported APIs or coding agents while using the gateway’s existing usage tracking, budgets and routing controls.

Sep 17, 21:49 UTC Resolved - Between 20:26 and 21:17 UTC on September 17, 2026, GitHub Copilot experienced degradation affecting several GPT models, including GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.6 Sol, GPT-5.3-Codex, and GPT-6 Astra.

The UN System Data Commons brings statistics from UN organizations into a connected platform built on Google’s Data Commons technology. Users can explore data through natural-language queries, thematic browsing and tools for AI research assistants. Google says UN experts validate the datasets, while advising readers to check underlying sources before citing important figures. Broader coverage planned for 2027 remains a future goal.

Vercel’s September Production Index reports that open-weight models processed 56% of AI Gateway tokens in August, while Anthropic accounted for 64% of estimated spending. A later update says GPT-6 Astra attracted more than twice Fable 5.1’s spending share over each model’s first twelve days. The figures describe gateway traffic, and spending uses published list prices rather than customers’ actual bills.

The rejectIfBusy option lets synchronous Workers AI inference requests fail when capacity is unavailable. Use it when your application should not wait in a capacity queue. Pass the option as the third argument to the Workers AI binding: const response = await env.AI.run( "@cf/google/gemma-4-26b-a4b-it", { messages: [{ role: "user", content: "Explain capacity queues." }], }, { rejectIfBusy: true }, ); const response = await env.AI.run( "@cf/google/gemma-4-26b-a4b-it", { messages: [{ role: "user", content: "Explain capacity queues."

Vercel added OpenAI’s GPT-Live 1 to AI Gateway for voice sessions that can listen and speak simultaneously. Applications can optionally delegate background work to a separately selected text model while conversation continues. The application remains responsible for permissions, confirmations and cancellation. Delegated requests are billed separately, and voice-session charges continue during that work; the integration requires AI SDK 7.

Mistral and Mozilla announced that Firefox Smart Window, currently in beta, uses Mistral models to assist browsing and interpret information from open tabs. Initial coverage includes France and North America, with the UK and Germany expected later in 2026. The companies say conversations are not stored on Mozilla’s servers by default and that Mistral agrees to zero data retention for this partnership.

TypeSafe AI’s Jev is available through Vercel AI Gateway for applications that need structured decisions rather than generated prose. It evaluates multiple questions in parallel and returns choices, scores or boolean probabilities. AI SDK 7 exposes it through an experimental evaluation API. The announcement recommends calibrating confidence against labeled workflow examples; callers can request Zero Data Retention and No Training options.

Google introduced Gemini 3.8 Live for responsive voice interaction and an Extended Thinking variant for more complex tasks. The models combine conversation with visual context and background tool calls; Google says Extended Thinking can reason while continuing to speak. Developer access is rolling out through the Gemini API and AI Studio, while Gemini Enterprise access remains a private preview with additional product availability planned.

Google describes a push beyond text translation toward AI that handles speech, cultural context and underrepresented languages. Its work includes community-built speech datasets and TranslateGemma, lightweight models trained across 55 languages that can run on devices without an internet connection. Supporting the world’s 1,000 most-spoken languages remains a goal, with local partnerships central to improving coverage.

Google highlights AI research in genomics, weather forecasting, disease detection and education. James Manyika points to openly available AlphaGenome Atlas predictions and WeatherNext 3 as recent examples, alongside research partnerships and workforce training. He argues that wider benefits depend on collaboration, access and responsible use, while acknowledging that economic gains are not automatic and that scientific applications carry misuse risks.

Google’s AI for Societal Impact collection organizes examples of research and community partnerships in disease detection, natural-disaster forecasting, learning and economic opportunity. The company presents these efforts as ways to turn scientific advances into practical uses. The collection’s introduction emphasizes working with local experts so that more communities can benefit from AI.

Google has opened an interactive version of its AI & Economy ATLAS for exploring AI use across countries and occupations. Research with Google DeepMind and MIT FutureTech surveyed more than 600 scientists in the United States and United Kingdom. Respondents reported saving almost seven hours weekly, but the study also identified validation work and experimentation bottlenecks that could delay gains in scientific discoveries.

Vercel added Google's Gemini 3.8 Live and Live Extended Thinking to AI Gateway for spoken conversations. Its announcement describes visual grounding, automatic switching among 97 languages and tool calls that can run while speech continues. Extended Thinking adds parallel multistep reasoning. Developers connect through the AI SDK realtime API using a short-lived token and a WebSocket connection.

You can now grant access to specific Workers and choose from four roles to control the level of access you give teammates, agents, and CI/CD workflows. Choose from four roles to control the level of access: Metadata Read-Only: View settings, metrics, logs, and traces without access to Worker code or the ability to make changes. Content Read-Only: Read Worker code, settings, and observability data without the ability to modify or deploy changes. Editor: Update and deploy a Worker without the ability to delete it. Admin: Everything in Editor, plus the ability to delete the Worker. Worker-level access controls are available today for all customers. You can configure them in the Cloudflare dashboard, through the API, or with Terraform. Roles designed for how teams build Give Metadata Read-Only to a debugging agent so it can inspect settings and observability data without seeing Worker code.

Vercel's AI SDK harness can now use native coding-agent subscriptions when the underlying agent supports them. Explicit provider credentials take priority in direct mode; automatic mode can use a subscription only without Gateway credentials. Gateway mode never reads subscriptions. Vercel says tokens remain managed on the host, with placeholder injection available only where the sandbox supports it.

Google's latest Dialogues on Technology and Society episode features NASA astronaut Christina Koch in conversation with James Manyika. The accompanying article says they discuss her International Space Station experience, the first all-female spacewalk and the Artemis II lunar journey. Their conversation also considers how astronauts work alongside robotics and AI, with Koch offering advice to people pursuing exploration careers.

Google says DevFest 2026 will run from October 1 through December 31, with more than 800 community events planned in 115 countries. Local Google Developer Groups will organize practical workshops and coding sessions covering tools including Gemini, Google Cloud and Firebase. This year's program emphasizes building AI applications, securing their deployment and scaling them, while local organizers tailor agendas to their communities.

AI Gateway can now require credentials for third-party provider requests. Credentials must accompany the request or be stored on the gateway. This setting prevents fallback to Unified Billing with Cloudflare-managed credentials. Turn on Require provider credentials in your gateway settings. To use the API, set byok_only to true in the request body of a PUT request to update the gateway: { "byok_only": true } To require provider credentials for one third-party request, set the cf-aig-no-wholesale header to true. This header cannot relax the gateway setting. Requests without applicable credentials then return an HTTP 400 response. Workers AI requests remain allowed, and the setting does not change their configured billing mode. For configuration details and request-level controls, refer to Prevent Unified Billing fallback for BYOK third-party providers.

Vercel's case study describes Tailscale's Aperture routing model requests through AI Gateway and running agents in Vercel Sandbox. Access follows tailnet identity instead of issuing each agent provider keys. Tailscale says this let it move from prototype to paying customers in months. The integration exposes request costs and can restrict routing to providers supporting zero data retention.

Vercel added an official GitHub Copilot adapter to the AI SDK harness layer. Applications can invoke Copilot through the same HarnessAgent interface used for other coding agents, reducing the application changes needed to switch agents. The adapter connects through the Agent Client Protocol using the harness ACP package.

Google highlights three ways its Search tools can support race preparation: creating a training schedule through AI Mode's Canvas, generating playlists after connecting a YouTube Music account, and comparing running equipment against personal requirements. The article describes shopping comparisons and local stock information as additional aids.

Mistral and Cloudera have announced a partnership to bring Mistral models into Cloudera's hybrid data platform. Mistral says the integration will support deployment in public and private clouds, on premises and in disconnected environments. The companies also aim to help enterprises build models using proprietary data while retaining control of their infrastructure, data and resulting models. The integration is planned.