# Morning digest: OpenAI promises to pay its own power bills at future data centers, AI giants buy coding-tool startups, and Cat Wu joins OpenAI

> OpenAI published a National Community Compact promising to pay the full cost of the power its future data centers use and to stop using NDAs with public officials.

Published 2026-11-05T06:05:00-08:00 · Daily Digest · Canonical: https://news.littlewolflabs.com/posts/2026-11-05-morning-digest

Morning digest for Thursday, Nov 5. OpenAI published a National Community Compact promising to pay the full cost of the power its future data centers use and to stop using NDAs with public officials. OpenAI and Meta each bought a startup that builds tools for AI coding agents, and Meta's purchase means Conductor's cloud service shuts down this month. Cat Wu, who led product for Claude Code, is joining OpenAI. Also today: cheaper status updates for Claude Code's background agents, a new open-weight coding model from Ricursive, voice mode and remote agents in Google's Antigravity, and Pangram's new AI image and video detector.

1\. **OpenAI says it will pay the full cost of the electricity and grid upgrades at future data centers it owns, and will stop using NDAs with public officials.** In a [post](https://openai.com/index/setting-new-standard-responsible-ai-infrastructure) published Wednesday evening, OpenAI released a National Community Compact, which "sets out what communities can expect from OpenAI at future campuses that we own and operate," and a Compute Campus Policy Blueprint with rule proposals for the whole industry. The compact says, "We'll pay the full cost of the electricity we use and the infrastructure needed to serve us," and "Before major spending begins, we'll provide financial assurances so other customers are protected if a project is delayed, reduced, or canceled." OpenAI also says it will "publish site-level water and electricity use," release independent noise studies, and "will not use non-disclosure agreements (NDAs) with public officials," apart from narrow exceptions such as security information and trade secrets. "Every future compute campus OpenAI owns and operates will have its own local compact," with a yearly public report. As an example of easing strain on the grid, OpenAI says its deal with Georgia Power for its planned Effingham County campus "allows the utility to reduce the power it delivers to us by up to 1,000 megawatts during periods of high demand." These are OpenAI's own promises; the compact covers only campuses OpenAI owns and operates, and the blueprint is a set of proposals, not law. OpenAI itself calls voluntary commitments "an important starting point." This follows our Nov. 4 post on [AI data centers as a midterm election issue](/posts/2026-11-04-midterms-ai-data-centers). *What to watch:* Whether these terms also show up at data centers that partners build for OpenAI, and what the first yearly water and electricity reports actually show.

2\. **OpenAI and Meta each bought a startup that builds tools for AI coding agents, and Conductor's cloud service shuts down Nov. 20.** OpenAI [says](https://openai.com/index/openai-to-acquire-pierre) it will acquire Pierre Computer Company, the maker of Code Storage, a service for hosting code, and of the Trees and Diffs libraries already used to browse files and review changes in the Codex app. "Customers can continue using Code Storage after the acquisition closes while we work to bring its features and functionality into Codex," OpenAI says, and Pierre's open-source projects will stay open source. The deal still needs regulatory approval. The same day, Meta acquired Conductor, a Mac app for running several AI coding agents in parallel, [Axios reports](https://www.axios.com/2026/11/04/meta-acquires-ai-coding-startup-conductor); terms weren't disclosed. Conductor co-founder Charlie Holtz [confirmed on X](https://x.com/charlieholtz/status/2118045232218743264) that the team is joining Meta to work with Nat Friedman and Alexandr Wang at Meta Superintelligence Labs, and said, "We have some big, crazy plans for the next interface." Conductor's [changelog](https://www.conductor.build/changelog) says "Conductor Cloud goes read-only on November 13 and shuts down on November 20," while the local Mac app "keeps getting new models and critical bug fixes until early 2027." *What this means for you:* If you use Conductor Cloud, move your workspaces before Nov. 13: right-click a cloud workspace and choose "Transfer to local," or download everything from Settings → Account → Export data.

3\. **Cat Wu, who led product for Claude Code at Anthropic, is joining OpenAI to lead product management for its core products.** "I'm joining OpenAI to lead product management for Core Products!" Wu [wrote on X](https://x.com/_catwu/status/2118137612577968402) Wednesday, adding, "we'll build agents that have your back and your team's, and keep you in control." OpenAI confirmed the hire, and Anthropic declined to comment, [Business Insider reports](https://www.businessinsider.com/cat-wu-joins-openai-from-anthropic-product-lead-coup-2026-11). An OpenAI representative told BI that she'll report to Tibo Sottiaux and work with Nick Turley on enterprise products and Ashley Alexander on consumer products. This follows Tuesday's post on [her leaving Anthropic](/posts/2026-11-03-cat-wu-leaves-anthropic). *What to watch:* Whether OpenAI's Codex and ChatGPT agent features start to pick up ideas that Claude Code users know well, now that the person who ran Claude Code's product side is at OpenAI.

4\. **Anthropic says status updates from Claude Code's background subagents now cost about 99% less of your usage.** Subagents are helper agents that Claude Code runs in the background on parts of a task. In a [post on X](https://x.com/ClaudeDevs/status/2118063143167693011) Wednesday, Anthropic's official ClaudeDevs account said, "A running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less." Haiku is Anthropic's smallest, cheapest model family. The 99% figure is Anthropic's own and applies to the status updates, not to the subagents' actual work. *When you'll feel it:* If you run long background subagents on a Claude plan, the status updates should now take a much smaller bite out of your usage limit. The work itself still costs the same.

5\. **Ricursive Intelligence released Ripple-65B Preview, an open-weight coding model it says beats much bigger U.S. open models on a coding benchmark.** Open-weight means anyone can download the model and run it themselves. The startup, which builds AI for chip design, [says](https://www.ricursive.com/ripple) the 65-billion-parameter model (parameters are the model's internal settings, a rough measure of size) was trained from scratch on 12 trillion tokens and handles up to 1 million tokens of context. It reports a score of "80.52" on TerminalBench 2.1, a test of coding agents working in a terminal, ahead of Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). It also says it built and trained the model "for only $32M," adding "$32M is our full R&D cost." Both are Ricursive's own numbers. The weights are on [Hugging Face](https://huggingface.co/ricursiveAI/Ripple-65B-Preview) under the Apache 2.0 license, and [OpenRouter](https://openrouter.ai/ricursiveai/ripple-65b-preview) lists it at $0.20 per million input tokens and $0.60 per million output tokens. Ricursive says it can run on local devices "such as NVIDIA DGX Spark." Early hands-on reports are mixed: one tester [on X](https://x.com/WescheNex1q/status/2118136461879087118) said that on the same DGX Spark and prompt, Qwen 27B worked on the first try, while Ripple took "3 rounds of self-fixes" and "still needed a 3-line patch." *What to weigh:* It's cheap and open, but the benchmark win is vendor-reported, so try it on your own tasks before you switch any coding work to it.

6\. **Google's Antigravity coding app is rolling out a voice mode and cloud-based "remote agents."** Version 2.29 of Antigravity, Google's agent-first coding app, adds a voice mode "powered by Gemini Live" that "lets you talk to Antigravity and hand requests to the agent by voice," according to the [changelog](https://antigravity.google/changelog). Google says "Voice mode is rolling out gradually, so you may not see it yet." Remote agents "are starting to roll out to Google AI Ultra and Pro subscribers," running conversations "in the cloud instead of on your computer," and you can also open them at antigravity.google.com/remote. The update also adds a built-in viewer for Jupyter notebooks (the .ipynb files data scientists use). *Try it:* If you're on Google AI Pro or Ultra, update Antigravity and look for the voice button in an empty message box and for cloud projects; if they're missing, the gradual rollout hasn't reached you yet.

7\. **Pangram, best known for detecting AI-written text, launched a detector for AI-generated images and video.** Pangram [says](https://www.pangram.com/blog/introducing-pangram-image-1-0) its new Pangram Image 1.0 model "achieves 99.96% accuracy" on its own test of 1,200 AI-made copies of 100 real photos, created with 12 image generators. It says a faster Flash 1.0 model, built for video, scores 98.7% on a deepfake video benchmark called DF26, "a 29-point leap over the next best detector." These are Pangram's own tests, not independent ones. You can now upload videos, which it checks frame by frame, the minimum image size drops from 512x512 to 256x256 pixels, and the detector is now in Pangram's Chrome extension and API. *What to weigh:* Detectors can get it wrong, including flagging real photos as AI-made, so treat a result as one clue, not proof, especially before you accuse anyone of faking something.