OpenAI and Anthropic public-post watch

Checked 14 September 2026 at approximately 3:10 am AEST (5:10 pm UTC on 13 September). Priority window: the preceding 24 hours. This is an incremental update to the 10 September bulletin.

Headlines

• Anthropic committed to embedded independent evaluators, and OpenAI said it will do the same — Anthropic CEO Dario Amodei proposed “pacing the frontier”: not halting training, but slowing capability gains enough for safeguards and third-party review to catch up. His three-step framework calls for permanent, employee-like access for external evaluators; common safety standards and limits among democratic-country labs; and eventually verifiable global coordination. Anthropic is committing unilaterally to the evaluator step. Sam Altman agreed that the frontier should be paced and said OpenAI will also provide independent evaluators with employee-like access, with details still to come. Amodei’s and Altman’s posts were just outside the strict window—about 27 hours and 24½ hours old respectively—but Anthropic researcher Sholto Douglas endorsed the proposals inside the window, arguing against both an unrestricted race and an absolute pause. These are significant voluntary commitments, but neither company has yet published operational details, timelines or an enforcement mechanism. Amodei’s post · Amodei’s essay · Altman’s response · Douglas’s in-window post

• OpenAI says it fixed three sources of degraded GPT-6 Astra behaviour and completed a usage reset — Tibo Sottiaux said some skills written for older models were triggering too often or blocking self-checking; an opt-in context-management experiment could stop early or answer an older message; and a small set of misconfigured engines caused measurable quality loss in long-tail traffic. OpenAI disabled the experiment, which it estimates affected roughly 4,000–5,000 users, removed the bad engines and made smaller follow-through and message-tracking improvements. Sottiaux later confirmed that the promised reset had fully propagated. These fixes were posted about 33–38 hours before this scan, so they are outside the priority window but new since the last bulletin; the claimed improvement is OpenAI’s assessment rather than an independent measurement. Quality update · Reset confirmation

• OpenAI brought the Git AI team in-house while promising to keep its agent-attribution tool open source — Sottiaux announced that Aidan Cunniffe and Sasha Varlamov are joining OpenAI. Their Git AI project tracks how coding agents contribute to a codebase; OpenAI says it will use the work to help businesses see where Codex is making a difference for individuals and teams. Romain Huet separately confirmed that Git AI will stay open source and continue receiving investment. No acquisition terms or product-integration timeline were disclosed. Sottiaux’s post · Huet’s post

• OpenAI says ChatGPT Sites has passed five million creations and added collaboration, private sharing and custom domains — Greg Brockman highlighted the milestone roughly three months after launch. The accompanying product thread says teammates can now jointly edit, save and publish a Site; creators can invite named people without making a Site public; deployment is twice as fast; ChatGPT can inspect a Site’s database; and eligible users can attach their own domains. The five-million figure counts sites created, not necessarily published sites or active users. Brockman’s post · ChatGPT’s product thread · Sites help page

• OpenAI exposed more of the Codex agent stack and GPT-Live voice architecture to developers — In public beta, the Agents API provides the managed harness behind Codex, including durable sessions, automatic context compaction, tool search, programmatic tool calling, subagents and a choice of OpenAI-hosted, partner or self-managed environments. OpenAI says there is no separate API surcharge beyond model and tool usage. GPT-Live-1 is now also available in the API as a $0.05-per-minute front-end voice layer that can listen and speak simultaneously, handle interruptions, and delegate reasoning or tools to Astra or another backend model. Both are catch-up items from 10–11 September rather than fresh 24-hour posts. Steve Coffey on Agents API · Sherwin Wu’s explanation · Agents API announcement · Wu on GPT-Live-1 · GPT-Live-1 announcement

Steve Yegge and scan notes

Steve Yegge’s newest main-feed post is from 11 September, outside the priority window. He asked what people use as an “agentic IDE” to multiplex 10–20 or more agents, naming several candidates and saying his current Emacs setup is not something he would recommend. It is a useful signal that multi-agent supervision remains a tooling problem for advanced users, but it is a question and personal workflow anecdote, not a product result. Yegge’s post

Across the representative OpenAI and Anthropic leadership, product, developer, research, safety and policy timelines checked, no new product launch or research result appeared inside the strict 24-hour window. Jack Clark’s only fresh main-feed post said this week’s Import AI newsletter would be skipped because of sudden movement on critical AI-development questions and family commitments; Kevin Weil shared an external essay about AI and mathematics. Neither was promoted to a headline because neither introduced a new company, product or research development.