OpenAI and Anthropic public-post watch

Checked 19 August 2026 at approximately 3:10 am AEST (5:10 pm UTC on 18 August). Priority window: the preceding 24 hours. The most recent recorded AI bulletin was 17 August, so one important later post outside the strict window is included and labelled.

Headlines

• OpenAI and NVIDIA are planning a 4.25-gigawatt AI campus in Ohio — Sam Altman said he was excited to work with NVIDIA after Jensen Huang published the terms of the PORTS-Pike project. NVIDIA says it will work with SB Energy to secure the land, power, and data-centre shell, with OpenAI as tenant and operator of an AI factory built on NVIDIA’s DSX platform. The initial deployment is expected to provide 4.25 GW, come online in phases from 2028 to 2030, and could accommodate roughly 1.5 million GPUs per hardware generation. NVIDIA says OpenAI’s existing and planned commitments amount to about 12 GW of NVIDIA compute through 2030, potentially rising to 16 GW if another 3.75 GW at the site is secured; it values that broader opportunity at roughly $600 billion. These are NVIDIA’s long-range estimates and contractual description, not delivered capacity or recognized revenue. NVIDIA also says its support covers defined lease, power, and residual-value obligations rather than the project’s entire cost. Altman’s post · Huang’s original article

OpenAI and Anthropic public-post watch

Checked 17 August 2026 at approximately 3:05 am AEST (5:05 pm UTC on 16 August). Priority window: the preceding 24 hours. Items covered in the 16 August bulletin were excluded.

Headlines

• OpenAI’s Tibo Sottiaux argues that cost per token is a misleading cross-model comparison — Sottiaux noted that tokenizers can split identical text differently, so a cheaper token need not mean a cheaper completed task. In one small sample spanning English, technical, multilingual, and numerical text, he said GPT-5.6 Sol’s tokenizer used 766 tokens versus an estimated 1,170 for Claude Opus 5, or about 34.5% fewer. His more useful conclusion is to compare price per successful outcome on your own workloads. Treat the sample as an OpenAI product lead’s illustrative comparison, not a general benchmark proving that Sol is always cheaper. Original post

OpenAI and Anthropic public-post watch

Checked 16 August 2026 at approximately 3:10 am AEST (5:10 pm UTC on 15 August). Priority window: the preceding 24 hours. Items already covered in the 15 August bulletin were excluded.

Headlines

• Codex multi-agent delegation can now choose any supported model, including Luna — OpenAI Codex developer-experience engineer Eric Provencher said the multi-agents v2 system can delegate work to any supported model, following reliability work on the feature. Greg Brockman amplified the update as progress toward a system in which users do not have to select models manually. The shipped capability is cross-model delegation; Brockman’s broader automatic-selection framing is directional and does not establish that a fully automatic model router has shipped. Provencher’s original post · Brockman’s post

Anthropic watch

• No new substantive employee post since the previous check — The public main-feed timelines reviewed did not surface a new Anthropic product, research, developer, policy, or company update after the 15 August bulletin. Jack Clark’s research-hillclimb post remains inside the rolling 24-hour window but was already covered yesterday and is not repeated here.

Steve Yegge watch

• No fresh visible post in the priority window — Steve Yegge’s public main-feed timeline still showed 11 August as its newest visible post, so no older item was recycled.

Coverage note

Public timelines were checked across notable OpenAI leadership, product, developer, research, safety, and policy staff including Sam Altman, Greg Brockman, Eric Provencher, Adam Fry, Tibo Sottiaux, Romain Huet, Nick Turley, Kevin Weil, Boaz Barak, Mark Chen, Michelle Pokrass, and Will Depue. Anthropic staff checked included Dario Amodei, Jack Clark, Boris Cherny, Amanda Askell, Alex Albert, Mike Krieger, Eric Jang, Jan Leike, and Chris Olah. Steve Yegge’s public profile was checked separately. Previously covered ChatGPT feature drops, Ultrafast mode, Claude content marking, EU AI Act guidance, GPT-5.6-Cyber and Daybreak access, usage resets, Claude Code diagnostics, prompt-injection benchmarks, and auto mode were not repeated.

OpenAI and Anthropic public-post watch

Checked 15 August 2026 at approximately 10:30 pm AEST (12:30 pm UTC). Priority window: the preceding 24 hours.

Headlines

• ChatGPT added quizzes, natural-language reservation search, Google Drive imports, and homepage suggestions — OpenAI’s Adam Fry described four additions in the week’s ChatGPT feature drop. Asking “Quiz me on [topic]” now generates an in-chat quiz; reservation search accepts ordinary descriptions of the desired venue; paid users can add Google Drive files to their ChatGPT Library and ask questions about them; and paid users will see suggested tasks on the home page. Greg Brockman and Tibo Sottiaux separately highlighted reservation search. The posts describe a product rollout, but they do not specify region-by-region availability or a complete list of eligible paid plans. Fry’s original post · Brockman’s post · Sottiaux’s post

OpenAI and Anthropic public-post watch

Checked 14 August 2026 at approximately 3:05 am AEST (5:05 pm UTC on 13 August). Priority window: the preceding eight hours.

Eight-hour headline

• No substantive employee-authored update surfaced inside the priority window — Public and indexed timelines across notable OpenAI and Anthropic leadership, product, developer, research, safety, and policy staff did not yield a new, verifiable product, research, developer, policy, or company update during the strict eight-hour window.

Since the previous run

• No additional high-signal employee post was verified — Broader checks back to the previous bulletin did not surface an employee-authored update important enough to displace yesterday’s coverage. Search indexes repeatedly presented older July posts with relative labels such as “hours ago”; those entries were excluded after their underlying publication dates were checked. Previously covered Claude content marking, OpenAI’s EU AI Act guidance, GPT-5.6-Cyber and Daybreak access, infrastructure policy, usage resets, Claude Code diagnostics, prompt-injection benchmarks, auto mode, GPT-Live, and model-harness interoperability were not repeated.

Coverage note

OpenAI staff checked included Sam Altman, Greg Brockman, Romain Huet, Nick Turley, Mark Chen, Boaz Barak, Jason Liu, Tibo Sottiaux, Vaibhav Srivastav, Atty Eleti, and Adam Fry. Anthropic staff checked included Dario Amodei, Jack Clark, Amanda Askell, Boris Cherny, Alex Albert, Thariq Shaukat, Jan Leike, Chris Olah, Mike Krieger, Cat Wu, and Eric Jang. X’s live search required sign-in, so candidate posts were cross-checked through public profile indexes and first-party pages; no unverified or stale item was promoted into the bulletin.

OpenAI and Anthropic public-post watch

Checked 13 August 2026 at approximately 3:05 am AEST (5:05 pm UTC on 12 August). Priority window: the preceding eight hours.

Eight-hour headline

• No substantive employee-authored update surfaced inside the priority window — The public employee timelines and indexed posts checked across OpenAI and Anthropic did not yield a new product, research, developer, policy, or company update in the strict window. The most important fresh signal was instead an Anthropic company guidance update on machine-readable marking, summarized below without attributing it to an individual employee.

Important company update in the priority window

• Anthropic will mark text and files produced by new Claude models worldwide — Anthropic says Claude models launched in the EU on or after 2 August 2026 will embed imperceptible watermarks in generated text and attach signed C2PA provenance metadata to supported files such as SVG, PNG, and JPEG. The policy applies across the API, Claude, Claude Code, Cowork, Claude Tag, and supported cloud partners wherever Claude is offered, not only in Europe. Existing models are still being brought into the scheme, and detection tooling is forthcoming. Anthropic cautions that a detected mark only indicates Claude may have processed content—not that Claude authored the underlying ideas—and that heavy editing, short passages, format conversion, or stripped metadata can make marks undetectable. Anthropic’s original guidance

Other important update since the previous run

• OpenAI consolidated its EU AI Act compliance guidance — OpenAI’s updated guidance links public training-data summaries for GPT-5.2, GPT-5.4 Nano, GPT-5.5, GPT-5.6 Luna, and GPT Image 2.0, and says its provenance stack combines C2PA credentials, SynthID watermarking for images and audio, a public verification portal, and a Content Provenance API. Unlike Anthropic’s new-model policy above, the page does not announce invisible watermarking for generated text. This update appeared after the previous bulletin’s check but outside today’s strict eight-hour window. OpenAI’s original guidance

Coverage note

Public posts and profile timelines were checked across OpenAI leadership, product, developer, research, safety, and policy staff including Sam Altman, Greg Brockman, Romain Huet, Kevin Weil, Nick Turley, Boaz Barak, Tibo Sottiaux, Vaibhav Srivastav, Atty Eleti, and Adam Fry. Anthropic staff checked included Boris Cherny, Thariq Shaukat, Jack Clark, Amanda Askell, Alex Albert, Jan Leike, Chris Olah, Mike Krieger, Cat Wu, and Eric Jang. Previously covered GPT-5.6-Cyber and Daybreak access, OpenAI’s Texas infrastructure letter, usage-limit resets, Claude Code diagnostics, prompt-injection benchmarks, auto mode, GPT-Live, and model-harness interoperability were not repeated.

OpenAI and Anthropic public-post watch

Checked 12 August 2026 at approximately 3:03 am AEST (5:03 pm UTC on 11 August). Priority window: the preceding eight hours.

Eight-hour headline

• No substantive employee update surfaced inside the priority window — The newest relevant public posts visible from the OpenAI and Anthropic employees checked were about 15 hours old. The items below are included because they appeared after the previous bulletin’s check and are the strongest product and company signals since then.

Other important posts since the previous run

• OpenAI introduced GPT-5.6-Cyber and broadened its Daybreak cybersecurity program — Greg Brockman described GPT-5.6-Cyber as a new model and said Daybreak was expanding to put frontier intelligence in defenders’ hands. Tibo Sottiaux added that the program now has Blue and Red access tiers and positioned the model for advanced defensive work. Sam Altman separately urged users to consider OpenAI’s models for defending their systems. These posts landed roughly 23 hours before this check, shortly after the previous bulletin was compiled. They describe access for authorized cybersecurity work, not unrestricted general availability. Brockman’s post · Sottiaux’s post · Altman’s post

OpenAI and Anthropic public-post watch

Checked 11 August 2026 at approximately 3:10 am AEST (5:10 pm UTC on 10 August). Priority window: the preceding eight hours.

Eight-hour headline

• Anthropic’s Claude Code lead shared practical usage-limit diagnostics — Responding to users who said they had exhausted Claude usage unusually quickly, Boris Cherny said Fable should be included with Teams Premium seats but not Standard seats. He recommended running /usage to identify token-heavy plugins and skills, and /effort to choose a lower reasoning-effort setting. In a separate reply, he suggested checking whether loops or routines were consuming the allowance and offered claude -p /usage as a way to collect a detailed report for debugging. This is troubleshooting guidance, not an announcement of higher limits or a general promise about how long a plan’s allowance should last. Plan and usage reply · Diagnostic follow-up

Other important post since the previous run

• An OpenAI researcher said the Hugging Face incident calls for cultural as well as technical change — Boaz Barak agreed with an external critique that the evaluation incident requires more than fixing individual issues. He said he was seeing hopeful signs and identified closer collaboration between security staff and researchers as a major part of the needed change. The post appeared roughly ten hours before this check, just outside the priority window. It is a notable internal perspective, but not the promised detailed postmortem or evidence that the broader cultural work is complete. Original post

Coverage note

Public timelines were checked across OpenAI leadership, product, developer, research, safety, and policy staff including Sam Altman, Greg Brockman, Romain Huet, Kevin Weil, Nick Turley, Noam Brown, Sebastien Bubeck, Boaz Barak, Jason Liu, Vaibhav Srivastav, Atty Eleti, Adam Fry, and Tibo Sottiaux. Anthropic staff checked included Boris Cherny, Thariq Shaukat, Jack Clark, Amanda Askell, Alex Albert, Jan Leike, Chris Olah, Mike Krieger, Cat Wu, and Eric Jang. No substantive OpenAI employee post was visible inside the strict eight-hour window. Previously covered ChatGPT Work examples, usage-limit resets, Claude Code auto mode, prompt-injection benchmarks, Astra, GPT-Live, and model-harness interoperability were not repeated.

OpenAI and Anthropic public-post watch

Checked 10 August 2026 at approximately 3:02 am AEST (5:02 pm UTC on 9 August). Priority window: the preceding eight hours.

Eight-hour headlines

• Codex understands an informal /loop instruction even though no such command exists — OpenAI’s Jason Liu said that sending the literal string /loop is enough for Codex to infer the intended behavior. This is an undocumented interaction pattern, not a newly shipped slash command, but it is a useful developer signal: natural-language intent can substitute for a formal command when users want the agent to continue iterating. Original post

OpenAI and Anthropic public-post watch

Checked 9 August 2026 at approximately 3:08 am AEST (5:08 pm UTC on 8 August). Priority window: the preceding eight hours.

Eight-hour headlines

• Boaz Barak argued for formal safety coordination without weakening competition — Responding to a suggestion that the two leading AI labs could simply meet and resolve the race between them, the OpenAI researcher said competition is beneficial, antitrust laws matter, and other contenders should not be discounted. He supports regulation and industry-wide mechanisms that prevent a safety race to the bottom, but said those arrangements should not emerge from private meetings between two companies. The post is a useful policy signal from inside OpenAI: cooperate through transparent, sector-wide safeguards while preserving competitive pressure. Original post