OpenAI and Anthropic public-post watch

Checked 17 September 2026 at approximately 3:15 am AEST (5:15 pm UTC on 16 September). Priority window: the preceding 24 hours. This is an incremental update to the 14 September bulletin.

Headlines

• Anthropic is merging Claude chat and Cowork, while putting Docs, Slides and Design directly inside conversations — Boris Cherny said users will be able to move between quick questions and longer delegated work without choosing a separate surface, with Claude carrying context across the work and continuing after the laptop closes. Claude Docs, Slides and Design can now be created and edited inside a conversation: users can directly change a document, comment on a slide or move a design element, then present, export or share the result. Existing Cowork chats, projects, artifacts, connectors and skills are meant to remain in place. The unified experience is rolling out gradually to Pro and Max over the next few weeks; the three creation tools are in beta on all paid plans. This was announced less than an hour before the scan, so availability will not be immediate or uniform. Cherny on the unified Claude experience · Cherny on Docs, Slides and Design · Claude’s launch thread · Anthropic’s launch post

• Steve Yegge says frontier-model cost—not local hardware—is now the binding constraint on his multi-agent “software factory” — In a new essay, Yegge says his Wheelhouse system peaked at roughly 25 Fable agents dispatching work to another 25 Opus and Sol agents, but that 21 Claude Max accounts no longer keep the system running continuously. He estimates current consumption would require about 55 accounts, or $12,000 per month, and says one account’s weekly Fable allowance now lasts only two to four hours. These figures, including his claimed 250–300 meaningful commits per day, are personal and self-reported rather than a controlled benchmark. The useful engineering lessons are broader: hand work off earlier in an agent’s context, shrink tasks and startup prompts, monitor prompt caching, route simpler work to cheaper models, and treat every persistent role as a fixed cost. He also describes cutting 650 mechanically enforced refusal sites down to 14 after accumulated “fences” made almost all work illegal—an unusually concrete account of governance controls paralyzing an agent system. Yegge’s post · full essay

• OpenRouter data suggests Astra led model spending last week while Luna led token volume — Greg Brockman amplified OpenRouter chief operating officer Peter Walker’s comparison of OpenAI and Anthropic usage, which says Astra was the top model by dollars spent and Luna the top model by tokens by a wide margin. That is a meaningful developer-adoption signal, but it describes traffic routed through OpenRouter—not total API, enterprise or consumer usage—and the post does not disclose the absolute values behind the shares. Brockman’s post · Walker’s original post

• OpenAI’s Data agent can analyze governed company data, build dashboards and act on approved findings — In a fresh post inside the priority window, Brockman highlighted ChatGPT Work operating across sources such as Power BI, Tableau, ClickHouse, Oracle Analytics and Redshift. The underlying product launched on 10 September: administrators install the Data plugin and relevant connectors, while existing table-, row- and column-level permissions continue to apply. It can investigate changes using the company’s own metric definitions, build shareable dashboards in existing BI tools, and—with approval—send findings or take actions through connected tools such as Slack and email. OpenAI says nearly all of its product organization and more than two-thirds of go-to-market staff use the internal capabilities behind it; that is an OpenAI-reported adoption figure, not an external evaluation. Brockman’s post · OpenAI’s product announcement · usage and permissions guide

• OpenAI now says it prepares explicit safety cases before major frontier reinforcement-learning runs — In an important post published after the previous bulletin but outside today’s 24-hour window, Sam Altman said OpenAI formulates a safety case in advance of RL runs expected to produce significant capability gains, extending safety review from completed-model deployment into development. He supports a consistent federal safety framework, independent auditors and shared standards for misalignment and monitoring, but argues labs can begin voluntarily without waiting for legislation or an antitrust exemption. His definition of “pacing” is slower development to pay for safety and monitoring—not stopping—and he says international coordination will eventually require government involvement. A follow-up framed both loss of human control and concentration of power in one company, person or country as outcomes to avoid. OpenAI has not yet published the safety-case template, decision thresholds or review authority, so the operating detail remains unclear. Altman’s safety-case and pacing post · follow-up on control and concentrated power

• OpenAI says it reassigned a quarter of production engineering to an internal AI-driven “defense factory” — In an a16z interview published after the previous scan, Greg Brockman said OpenAI paused those engineers’ existing projects and used its own models to search its infrastructure for vulnerabilities. He says the team fixed serious findings until Astra stopped discovering additional critical issues within its current capability, and is automating the loop from discovery through triage, remediation, deployment and validation. Saturating one model’s findings is not evidence that the systems are vulnerability-free; Brockman acknowledged that each stronger model will require another pass, and the claims have not been independently verified. The same interview disclosed that OpenAI cancelled Sora this year to focus on its agentic-coding cycle, and described Astra’s 24-hour computer-use runs as enough for his personal “AGI era” label while conceding that its capabilities remain jagged. original a16z interview

Scan notes

Sam Altman and Tibo Sottiaux teased a high-volume week of OpenAI releases, but did not name a product or provide technical details, so the posts are not treated as announcements. Jack Clark’s fresh Anthropic post was a light comment on the new DeepMind Institute rather than an Anthropic development. Across the representative leadership, product, developer, research, safety and policy timelines checked, the unified Claude launch was the only new first-party product release inside the strict window.