A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
A model that solves engineering tasks end to end costs $0.75 per million input tokens
Gemini 3.8 Flash lands at $0.75 per million input tokens and $3.75 per million output, an introductory rate that expires on 31 December 2026 before doubling to $1.50 and $7.50. Google reports it outperforms most larger frontier models on the DeepSWE v1.1 benchmark at autonomously solving complex engineering problems end to end, and it is available through AI Studio, the Gemini app, AI Mode in Search and Gemini Enterprise. A separate variant, 3.8 Flash Cyber, finds vulnerabilities and writes patches for them, with the Chrome Security team measuring 2.6 times more correct patches than far larger commercial models, but it is reachable only through a vetting programme called Fairwind.
Your self-hosted agent workers now scale as a pool and drive a desktop
Cursor's self-hosted workers now organise into team pools, named queues that expand with demand, hibernate when idle, and let any free worker claim a request rather than being pinned to one repository. Those workers also gained computer use on Linux and Mac, so an agent can click, type, take screenshots and drive a browser on your own hardware. Code, build outputs and secrets stay inside your network while inference and planning stay in Cursor's cloud, across AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel and E2B.
The MCP servers your org runs can be pushed to every developer's Claude Code
Claude Code 2.1.259 adds managedMcpServers, letting an organisation supply HTTP and SSE MCP servers to its users instead of each developer wiring them up by hand. The release also adds a permission-prompts none option for unattended headless hosts and a JSON flag for plugin validation, and it recognises GitLab merge requests. A security fix extends the Bash Read deny rules to cover files named in option values, which they previously missed.
Registering a few JavaScript tools gives an agent your app's actions instead of its buttons
Site tools are ChatGPT's implementation of the WebMCP standard: a page registers tools in JavaScript through document.modelContext.registerTool, and the agent then works the same live page and signed-in session the user is on. Codex and ChatGPT Work can discover and call them inside the desktop app's built-in browser, and Codex will write the integration against your app's existing logic and permissions. The documentation is explicit that tool definitions and results are untrusted, that tools inside iframes are not discoverable, and that the declarative HTML-forms API is not yet supported.
Ask an assistant which tool to use and the answer often traces to manufactured sites
Researchers put identical top-five product prompts to Perplexity's sonar and sonar-pro models across 380 software categories on 2 September, logging 7,534 citations across 2,055 distinct domains. Nearly 60% of the cited sources ranked outside the top 100,000 sites on the Tranco list, and 23.4% fell outside the top million. Three domains registered between December 2023 and May 2024 shared nameservers and page templates and had generated 215,128 machine-made pages of best-software buying guides between them.
