OpenAI and Anthropic public-post watch

Checked 22 August 2026 at approximately 3:10 am AEST (5:10 pm UTC on 21 August). Priority window: the preceding 24 hours. Because the previous recorded bulletin was checked on 19 August, two major OpenAI posts published shortly after that check are included separately.

Headlines

• OpenAI says Codex has passed 20 million active users and is giving paid users a banked reset — Tibo Sottiaux said Codex crossed 20 million active users during the week and that every Codex and ChatGPT Work user would receive a banked usage reset that can be redeemed later. He separately said OpenAI is still investigating reports of unexpectedly fast usage depletion and has not found a general abnormality so far. One specific pattern did involve unsupported sub2api traffic being flagged by fraud systems; Sottiaux stressed that official clients and open-source clients using Sign in with ChatGPT remain supported. The reset is a one-off credit, not an announced permanent quota increase, and the broader usage investigation remains open. 20-million-user and reset post · Usage-limit clarification

• ChatGPT added an Apple Messages plugin on Mac — Romain Huet said ChatGPT can now read, draft, and send iMessages, as well as answer questions about frequent contacts, conversation themes, and missed follow-ups. The underlying announcement says the plugin can search and summarize conversations and is available in ChatGPT Work and Codex on desktop. This describes available actions after the user connects and authorizes the plugin; it does not imply that messages are sent autonomously by default. Huet’s post · ChatGPT announcement

• GPT-Image-2 can now generate transparent backgrounds — Sottiaux highlighted transparent-image generation in both ChatGPT and the API. OpenAI Developers calls the API capability a preview and positions it for reusable product images, design assets, mockups, and marketing graphics that can be placed on arbitrary backgrounds. Sottiaux’s post · Developer announcement

• Anthropic is preparing a no-retention data arrangement for Mythos-class models — Claude Code creator Boris Cherny said the most advanced model class needs additional safety measures while enterprise customers need privacy and compliance controls. His proposed balance lets customers own and control their data while Anthropic retains none, and he said it is coming “this fall.” This is an employee description of a forthcoming approach, not a generally available feature today; the post does not spell out its technical architecture or eligibility. Cherny’s post

• OpenAI’s first NVIDIA Vera Rubin racks are running its training stack — Greg Brockman called the deployment a major partnership milestone, amplifying OpenAI infrastructure leader Uday Ruddarraju’s announcement that the company’s first Vera Rubin racks had arrived and were already running the training stack for next-generation frontier pre-training. The post confirms operational integration of the first racks, not completion of the much larger planned capacity discussed in the prior bulletin. Brockman’s post · Ruddarraju’s original post

Major updates since the previous check

• OpenAI paused some frontier reinforcement-learning training on safety grounds — Sam Altman said OpenAI paused some frontier RL training because capabilities were beginning to outpace its alignment, security, and monitoring standards. He framed the move as unilateral action pending broader shared safety standards and said safety confidence will increasingly set the pace of progress. A follow-up clarified that OpenAI still expects to ship strong models soon and that the pause affects further-out releases. This is a pause in some training, not a halt to all model development. The post appeared less than two hours after the previous bulletin’s cutoff. Altman’s post and follow-up

• OpenAI previewed Private Safety Processing while retaining Zero Data Retention for frontier models — Greg Brockman said the company is pursuing technical and policy approaches intended to improve both business privacy and safety. The quoted announcement says Private Safety Processing is designed to identify risks across related interactions without giving OpenAI personnel access to the underlying content, while continuing to offer Zero Data Retention for frontier models. It is currently described as a preview, so the post should not be read as proof of universal deployment or as a complete specification of the privacy guarantees. Brockman’s post · OpenAI announcement

Steve Yegge watch

• No substantive fresh main-feed update in the priority window — Yegge’s newest visible post, from 20 August, is the third comic in his Fable-agent series: “My Fable agents cannot physically keep a secret.” It is newer than the prior bulletin but does not communicate a concrete product, research, policy, or company change, so it is not promoted to the headlines. Original post

Coverage note

Public main-feed timelines were checked across notable OpenAI leadership, product, developer, research, and infrastructure staff, including Sam Altman, Greg Brockman, Romain Huet, Tibo Sottiaux, Mark Chen, Nick Turley, and Uday Ruddarraju. Anthropic checks included Dario Amodei, Boris Cherny, Alex Albert, Jack Clark, Amanda Askell, and Jan Leike. Steve Yegge’s profile was reviewed separately. Previously covered OpenAI–NVIDIA capacity plans, Claude Code /design and CPU improvements, the optional one-million-token Codex context, token-price comparisons, and Dario Amodei’s regulation and biology thread were not repeated.