OpenAI and Anthropic public-post watch

Checked 26 August 2026 at approximately 3:10 am AEST (5:10 pm UTC on 25 August). Priority window: the preceding 24 hours. Important employee posts published after the previous check but outside that window are included separately.

Headlines

• OpenAI published the first measured results for Jalapeño, its first custom inference chip — Greg Brockman praised the team and linked OpenAI’s new engineering report. OpenAI says that, across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T on the public InferenceX benchmark, Jalapeño delivered 1.5–1.9 times more peak-throughput work per watt and 1.7–3.6 times lower end-to-end latency than the comparison systems. It also says the chip was taped out in nine months with AI assisting design and verification, and that deployment in OpenAI infrastructure is planned by year-end. These are OpenAI’s first reported measurements, not independent production-scale validation; qualification and broader model testing are still under way. Brockman’s post · OpenAI’s detailed results

• The five-hour limit is returning for Plus accounts across ChatGPT Work and Codex — Tibo Sottiaux said OpenAI planned to restore the limit on 26 August after previously postponing it. He gave two reasons: smoothing compute demand so weekly usage can remain generous, and preventing relatively casual Plus users from accidentally exhausting a whole week’s allowance. Sottiaux said the five-hour limit will remain disabled for the $100 and $200 Pro subscriptions for the coming months. This is a rolling-window restriction for Plus, not a stated cut to the separate weekly allowance; the post does not give the exact rollout hour. Sottiaux’s post

• Steve Yegge argues that large agent organizations will need “law,” not just sandboxes — Drawing on his self-reported Wheelhouse setup of roughly 50–60 agents, Yegge says his agents converged on a governance system of constitutions, offices, precedents, rule lifecycles, and mechanically enforced “fences.” His core distinction is that a fence refuses an unauthorized action according to explicit organizational policy, while a sandbox tries to contain an agent. He forecasts that cheap Fable-class models will bring hundreds or thousands of AI workers into companies and force institutional knowledge to become explicit and machine-operable. This is an anecdotal design thesis and forecast based on one unusually intensive personal deployment, not a demonstrated general result. Yegge’s original post · Full essay

Important updates since the previous check

• OpenAI says the promised Codex reset and several usage fixes landed — Sottiaux said the reset had propagated to accounts and fixes had shipped for the usage problems identified a day earlier, including image inefficiencies in long multiply-compacted sessions, high-tail Computer History usage, and excess title-generation usage. He said users should notice an improvement and that more changes were coming, so this confirms delivery of the initial work rather than closure of the broader usage investigation. Sottiaux’s confirmation

• A GPT-Live research lead described full-duplex voice as a sustained high-risk research bet — Kundan Kumar said OpenAI leadership continued supporting full-duplex models even when success appeared deeply improbable, because it was unclear that the approach could not work and incremental understanding kept making the problem more tractable. Greg Brockman amplified the post as an example of OpenAI’s ability to make long-term research bets. This is useful context on how GPT-Live was developed, not a new model release or performance claim. Kumar’s post · Brockman’s amplification

Anthropic watch

• No substantive fresh Anthropic employee post was visible in the priority window — Recent main-feed activity from the Anthropic leaders, product staff, researchers, and Claude Code staff checked was either older than the window or unrelated to AI product, research, developer, policy, or company developments. Previously covered posts about Mythos-class no-retention arrangements and Claude Code performance were not repeated.

Coverage note

Live public main-feed timelines were checked across representative OpenAI leadership, product, developer, research, and infrastructure staff, including Sam Altman, Greg Brockman, Tibo Sottiaux, Romain Huet, Nick Turley, Mark Chen, and Kundan Kumar. Anthropic checks included Dario Amodei, Jack Clark, Boris Cherny, Alex Albert, Amanda Askell, and Mike Krieger. Steve Yegge’s profile and new essay were reviewed separately. Previously covered Codex usage diagnoses, pricing changes, privacy announcements, the Apple Messages integration, image updates, and frontier-training pause were not repeated.