Know
What changedCursor agents can build and publish web apps
Cursor’s cloud agent can scaffold an app, provision Postgres, configure Cursor SSO, request secrets, run security checks, and deploy to a shareable URL after you explicitly publish. It works with new or existing Cursor Origin and GitHub repositories; GitLab support is planned. Paid plans with cloud agents are required. Team admins can review apps and secrets, set rules, approve blocked publishing, and unpublish.
Try it for prototypes, but review secrets and security checks before publishing.
Mistral Studio adds managed agents and workflows
Mistral Studio now offers managed agents with skills, sandboxes, subagents, hooks, and a durable runtime. Workflows combine rules, agent reasoning, human approvals, and event triggers, with generated code available for inspection. Apps include hosting, authentication, and access controls. Deployments can run in Mistral’s cloud, a customer VPC, on-premises, or mixed environments; private-cloud and on-premises options were described as planned for late November.
Evaluate Studio if you need managed agent workflows across cloud and private infrastructure.
Claude Enterprise adds generally available inference hooks
Anthropic made prompt and pre-tool-call inference hooks generally available for Claude Enterprise. They can send prompts and tool calls to an organization’s DLP or security tools for allow-or-deny decisions before inference or tool execution. Other additions include a beta security center, centralized admin management and access controls, plus a beta EDR integration for the local Cowork/unified Claude VM.
If you manage Enterprise, test hooks against your existing policies before broad rollout.
Gemini Omni 1.2 Flash lowers video-generation pricing
Google says Gemini Omni 1.2 Flash is generally available through the Gemini API and Gemini for business. It generates or edits clips up to 15 seconds, with one-second duration choices, conversational edits, and improved adherence to references. 720p output costs $0.07 per second, 30% below Omni 1.1 Flash; drafts can start at 360p or 720p and upscale to 1080p or 4K.
Recheck video-generation budgets and test whether iterative edits fit your workflow.
Inkling 2 Small is available through Tinker
Thinking Machines Lab introduced Inkling 2 Small, a 594-billion-parameter mixture-of-experts model with about 17 billion active parameters, image input, and variable effort. It is available with closed weights through Tinker. The company says it scores 31 on the Artificial Analysis Intelligence Index and plans to release full weights under Apache 2.0 after external safety testing, targeted for the following week.
Use Tinker to evaluate it now; treat the open-weight release timing as a plan, not a certainty.
Baidu releases ERNIE 5.5 Flash Base checkpoint
Baidu released an openly available pretraining checkpoint with about 280 billion total parameters and 14 billion active per token, published in FP8 and BF16 formats. It is intended for fine-tuning and research, rather than being a ready-made instruction model. Contemporary coverage notes geographic license exclusions and commercial-use authorization thresholds, so check the terms before adopting it.
Review the license and hardware needs before planning a fine-tuning experiment.
OpenAI opens limited access to GPT-6 Cyber
OpenAI introduced GPT-6 Cyber for scoped vulnerability research, with API and Codex access initially limited to select approved Daybreak Red customers. OpenAI reports a 48.5% ExploitGym score, compared with 32.1% for GPT-5.6 Cyber, and classifies its cyber capabilities as Critical under its Preparedness Framework, up from High. This is a restricted release, not general developer availability.
If you work in an approved security program, check eligibility and keep testing within authorized scope.
Perplexity launches persistent Teammates agents
Perplexity launched Teammates, agents designed to handle ongoing work, monitor changes, coordinate with other agents, and reuse projects, skills, memory, and permissions from Perplexity Computer. Users can inspect or redirect their work. The initial web release is for Max subscribers; mobile apps are planned, with no availability date supplied in the captured summary.
Consider it for monitored recurring tasks if you already use Perplexity Computer.