OpenAI and Anthropic public-post watch

Checked 10 September 2026 at approximately 3:10 am AEST (5:10 pm UTC on 9 September). Priority window: the preceding 24 hours. This is an incremental update to the 7 September bulletin.

Headlines

• Anthropic released an interactive model for exploring how AI could reshape the US economy by 2030 — Co-founder Jack Clark said the tool varies assumptions about AI capability and diffusion, and is intended to help the public and policymakers prepare rather than predict a single outcome. The model treats occupations as bundles of tasks that AI may augment, automate, leave unchanged or create; its modest, substantial and extreme scenarios all increase total output, while the more transformative cases automate more knowledge work and shift more gains away from labour. Users can set their own assumptions and compare them with a survey of more than 10,000 Americans. These are deliberately simplified scenarios, not forecasts, and they do not cover every possible constraint or outcome. Clark’s post · Clark’s follow-up · Anthropic’s explorer

• OpenAI may temporarily stop taking new Pro subscriptions if demand for GPT-6 Astra keeps growing at its current rate — Tibo Sottiaux described demand as unprecedented even relative to OpenAI’s previous steep growth, said the team was using every available lever to sustain it, and put service quality for existing users ahead of adding new subscribers. Sam Altman endorsed that priority while acknowledging that pausing sign-ups would be undesirable. This is a capacity warning, not an announced pause or a change to existing subscription entitlements. Sottiaux’s post · Altman’s response

• An Anthropic pretraining researcher resigned over the frontier-model race, and Anthropic’s alignment-science lead publicly agreed with the core warning — Jacob Coxon, who said he spent the past three years doing pretraining research at OpenAI and Anthropic, wrote that neither company was acting responsibly and called the race toward self-improving systems a gamble that should not be decided inside private companies. Evan Hubinger replied that he personally assigns a greater-than-10% chance to AI killing all humans within the next decade, and said Anthropic does not yet have a plan to align superintelligence or a clear path to one. These are the researchers’ risk assessments and forecasts, not measured probabilities or evidence that present systems can cause such an outcome; Anthropic had not published a company response at check time. Coxon’s resignation thread · Hubinger’s response

• OpenAI says a roughly 10,000-agent system produced a solution to the Navier–Stokes Millennium Prize Problem using a still-training model more capable than GPT-6 Astra — Greg Brockman called the result a milestone for AI, mathematics and scientific discovery. OpenAI says the agents found a finite-time singularity construction after about 88 hours, with Astra taking another 17 hours to formalise and verify it in Lean; the Navier–Stokes portion used about 2.7 million agent messages and 130 billion output tokens. The company published the paper and formal proof, but said it does not intend to claim the prize. The result was not independently validated during this scan, so “solution” remains OpenAI’s claim rather than settled mathematical consensus. OpenAI also acknowledged that rumours of related work by Anthropic employee Levent Alpöge and NYU mathematician Tristan Buckmaster prompted the evaluation, while stating that its researchers and agents did not see their work before public release. Brockman’s post · OpenAI’s report, paper and Lean proof

• ChatGPT Images 2.5 is rolling out across all tiers, with two new API models — Sam Altman announced the release, which OpenAI says improves reference-image fidelity, targeted and multi-turn edits, lighting, textures and complex layouts while reducing generation latency by up to 50% versus Images 2.0. ChatGPT adds Sketch, templates, on-image comments and prompt sharing; Flare is the faster default API model and Sunburst is the slower, higher-precision option. The rollout covers ChatGPT, ChatGPT Work and Codex on desktop, mobile and web, while existing ChatGPT image limits remain unchanged. The performance and quality comparisons are OpenAI’s own. Altman’s post · OpenAI’s announcement · ChatGPT release notes

• OpenAI highlighted a 16-plugin collection aimed at small businesses — Greg Brockman amplified the collection as a way to take routine work off business owners’ plates. This is a curated discovery surface rather than a new plugin architecture or model release, but it is a useful product signal about OpenAI’s push to package connected-agent workflows by audience. Availability and data access still depend on the individual plugins, plans and connected services. Brockman’s post · OpenAI’s original post

Just outside the priority window

• Anthropic’s Boris Cherny publicly compared GPT-6 Astra’s prompt-injection risk with Claude and Gemini models — Shortly before the 24-hour cutoff, Cherny said Astra appeared roughly on par with Gemini Flash and Opus 4.8 on the comparison he shared, and argued that evaluating and naming competitors encourages labs to train safer models. Tibo Sottiaux responded that Astra is OpenAI’s most aligned model to date. Neither short post establishes that prompt injection is solved in general: the claims concern particular evaluations, and attack success varies materially with the benchmark, tools and threat model. Cherny’s post · Sottiaux’s response

Steve Yegge watch

• No new substantive main-feed post appeared inside the 24-hour window — Yegge’s newest visible main-feed post was from 7 September. It updated the same Wheelhouse maintainability story covered in the previous bulletin: Fable 5.1 removed roughly 30% of the agent-built system after Fable 5 had driven it into an unmaintainable state, and Yegge reported that the resulting system was faster. This remains an anecdote from one bespoke multi-agent codebase, not a controlled model comparison. Yegge’s latest post

Coverage note

Live public timelines were checked for representative OpenAI leadership, research, product and developer staff, including Sam Altman, Greg Brockman, Tibo Sottiaux, Romain Huet, Nick Turley and Jason Liu. Anthropic checks included Dario Amodei, Jack Clark, Boris Cherny, Evan Hubinger, Alex Albert, Cat Wu and Thariq Shihipar, plus Jacob Coxon’s resignation thread. Steve Yegge’s main feed was reviewed separately. The Anthropic economic explorer, Astra-capacity warning, Coxon and Hubinger safety posts, Navier–Stokes claim, Images 2.5 release and small-business plugin collection fall inside the strict window; Cherny’s prompt-injection comparison is included separately because it was published about 28 minutes before the cutoff and had not appeared in the prior bulletin.