AI Product Updates: New Features and Releases

Templafy released an MCP server that lets ChatGPT, Claude, and Copilot generate PowerPoint decks that follow brand guidelines.

OpenAI is piloting Sponsored Agents in ChatGPT Ads, letting users open a clearly labeled, advertiser-sponsored conversation and then head to the merchant site, alongside Ads Manager updates with HubSpot and Shopify integrations, tested with select US advertisers.

The open-source, Mac-native coding agent Droppy Code shipped 9 releases in 30 hours, with Hydra now running 20 agents in parallel, Escape to stop and rewind a turn, new shortcuts, and support for Pi and Z.ai as providers.

Andrew Yang claims AI agents have polluted the real web, so OpenAI and Anthropic may need synthetic internet environments to train their bots, which he suggests could explain talk of slowing down.

A Zhihu Frontier livestream with Bingxiang He, "Rethinking OPD: One Query, Hundreds of Steps," covers LLM training, distillation, and RL/OPD.

AgentLayer now supports the Arc network, letting agents explore Arc's agent economy ecosystem.

tldraw thanks attendees of its flash launch event; flash is a free timeline animation workspace for drawing shapes, recording drag motions, and editing and exporting lo-fi hand-drawn animation.

Nomads.com now supports importing flights from Flighty: export the CSV in the app, upload it on the Nomads settings page, and trips appear on your profile immediately.

Ripple released a developer kit that integrates XRP and RLUSD payments into Stripe and Tempo's AI agent standard, letting agents make recurring payments and manage wallets across blockchains.

Lovable's CEO says internal apps built on Lovable by its own teams rose from 11 in April to more than 80 by August across 17 departments, with about 30 from the People team alone.

TestingCatalog's daily brief covers Claude merging chat and Cowork for Pro and Max, paid beta for Docs, Slides and Design, OpenAI's main release slipping a week, Grok Bot's 1Password support, and Muse invite codes.

Zhipu says GLM-5.3-Flash production inference now runs on over 100,000 domestic AI accelerators, with end-to-end throughput up 3.2× in under two weeks, much of it optimized by a GLM-5.3-driven agent.