OpenAI published the first measured results for Jalapeño, the custom chip it designed with Broadcom to run — not train — AI models. On InferenceX, an independent public benchmark, Jalapeño produced 1.5x to 1.9x more AI work per watt of electricity and answered 1.7x to 3.6x faster end-to-end than the commercial systems it was compared against, including Nvidia’s GB300. On chatty, back-and-forth workloads the gap widened to as much as 4.1x. In plain terms: OpenAI says it can now serve answers faster while using far less power, which is the single biggest cost line in running a chatbot at global scale. The caveat matters for anyone writing about it — OpenAI says deployment volumes will be “very small” through the end of 2026 and it will keep buying Nvidia hardware. This is a signal of direction, not a switch being flipped.
Ahead of a listing that could rival SpaceX’s for size, Anthropic is preparing to pitch investors on a “total addressable market” of more than $30 trillion — roughly the size of the entire US economy, and larger than the $28.5 trillion SpaceX used. Around $23 trillion of it is attributed to enterprise work that AI could eventually perform. Two pieces of context worth holding onto: a total addressable market number assumes a company captures 100% of a market, which never happens, and Anthropic’s own revenue projection for 2028 is $190–200 billion — about 0.65% of the figure. TAM slides are a genre of sales document, not a forecast. Expect this number to be quoted at you out of context for months.
Hugging Face — the platform where most of the world’s open AI models are shared, now hosting over three million models and a million datasets — has retained a bank to gauge acquirer interest at a valuation of $13 billion or more, nearly triple its 2023 price. No buyer has been named and no deal has been struck. CEO Clem Delangue has publicly said the company is close to profitability and optimising for long-term sustainability, which has left observers unsure how committed the sale process really is. Separately, OpenAI published its own account of a security incident involving Hugging Face and the steps it is taking on model security and monitoring — a reminder that the open-model supply chain is now infrastructure, with all the risk that implies.
Chris Malone, OpenAI’s head of data centres, has left the company — the fourth senior departure in the run-up to a planned 2027 listing, following chief revenue officer Denise Dresser, chief operating officer Brad Lightcap and Fidji Simo, Sam Altman’s second-in-command. Roughly 13 executives have left OpenAI this year. Malone joined in March 2025 to build out Stargate, the data-centre venture with Oracle and SoftBank, after infrastructure roles at Meta and Google. OpenAI says it “recently reorganized our infrastructure organization” and has a strong team in place. From a communications standpoint this is a textbook case of a departure story becoming a valuation story: OpenAI filed its prospectus confidentially in June and is defending an $852 billion valuation.
An internal Tencent piece attributed to Tang Daosheng and dated 26 August argues that the competitive question in AI has shifted from who moved first to who can sustain investment and iteration over years. It lands alongside two data points: Hy3 is now released and available globally with Hy4 described as still an expectation rather than a result, and WorkBuddy recorded roughly 20.97 million visits in Q2 2026 — first among the seventeen leading native AI desktop office agents tracked by Analysys. See the Tencent Mentions section below for the full company watch.
Codex cuts prices and retires two modelsGPT-5.6 Sol pricing drops more than 20% for three months across the API, ChatGPT Work and Codex. GPT-5.4 and GPT-5.4 mini retire from Codex on 31 August.
ChatGPT for Teachers expands to 55 US school systemsMore than 100,000 additional educators and staff get access, along with training and support. One of the largest single institutional AI rollouts announced this year.
Anthropic’s invisible watermark now rides on everything Claude writesAn imperceptible marker is woven into text from new Claude models — including Claude Code and Cowork — and survives copy-paste. Driven by Article 50 of the EU AI Act but applied worldwide. A detection API is coming.
The dedicated Substack tools were not available to this session, so newsletter coverage below was assembled from the same publications as they arrive in your Inoreader “AI” folder — including Latent Space, Ben’s Bites, The Neuron, AI Valley, AI Breakfast, Exponential View and The AI-Augmented Engineer. Paid-tier flags could not be verified without the Substack connector and have been omitted rather than guessed.
Argues the next competitive line for software companies is not a better interface for humans but a machine-readable surface an AI agent can operate directly.
Frames the shift as structural: software that only exposes buttons and screens becomes invisible to the agents increasingly doing the clicking.
Practical read-across for any organisation with a website or newsroom — being legible to agents is starting to look like being findable in search did twenty years ago.
Leads on Anthropic’s total addressable market pitch and does the arithmetic the slide leaves out — the company’s own 2028 revenue forecast is about 0.65% of the number.
Also covers OpenAI’s Jalapeño chip results and a new AI-ready Mac mini from Apple.
The Neuron’s framing — that TAM decks exist to justify data-centre spending before revenue arrives — is the most quotable line of the day.
Straight coverage of the Jalapeño benchmark results, with Anthropic’s $30 trillion claim as the secondary item.
Earlier in the week the same newsletter covered Nvidia sending AI chips into orbit and humanoid robots breaking a human record — a run of “AI moves into the physical world” stories worth tracking as a theme.
Reports that OpenAI has completed a model codenamed Bel intended to sit at the centre of the GPT-6 release. Not confirmed by OpenAI — treat as reporting, not announcement.
Follows the same newsletter’s coverage of “Ox Alpha,” an unattributed stealth model that has been winning over developers without anyone claiming ownership.
Harvard is selling a bootcamp fronted by AI avatars of its own faculty — a live test of how far an institution will lend its brand to a synthetic presenter.
Also flags that OpenAI is testing a web standard for agent-ready sites, and MIT research finding chatbot assistance can weaken a person’s judgment when the chatbot is taken away.
Covers OpenAI’s new Admin plugin for ChatGPT Work and Codex as an enterprise governance story rather than a feature story.
Second item explains Anthropic’s invisible text watermark in plain terms: “You will not see the mark, but software may” — and what that means for corporate content review.
Leads with an OpenAI capability moving to the free tier, paired with a sceptical piece on who is genuinely winning and losing in the AI race.
Recent issues from the same newsletter have covered how frontier AI companies score on emergency safety preparedness, and LinkedIn’s “Ew, That’s AI” button passing a million clicks in three weeks.
A single long prompt that turns subject-matter expertise into a curriculum outline, launch plan and distribution strategy.
Adaptable for internal training material or a podcast series structure rather than a commercial course.
From Your AI Feeds
Inoreader AI Folder.
15 distinct articles in the last 24 hours from your “AI” folder, after collapsing the duplicate copies that arrive via both the individual feeds and the “Newsletters on AI” aggregate. Coverage is dominated by two stories: OpenAI’s Jalapeño chip results and Anthropic’s $30 trillion IPO pitch. A promotional post from Pivot to AI has been excluded.
OpenAI has launched a new blog dedicated to how transformative AI could reshape power, governance, the economy and individual freedom. It is a notable move for a company weeks away from public-market scrutiny: rather than confining itself to product news, OpenAI is staking out ground in the political argument about AI.
OpenAI’s account of a security incident involving Hugging Face, and the changes it is making to model security, monitoring and alignment as a result. Timing is everything here — it lands the same week Hugging Face is reported to be shopping itself at $13 billion.
Reporting that OpenAI has completed a model codenamed Bel expected to anchor the GPT-6 release. This has not been confirmed by OpenAI and should be handled as unattributed reporting if it comes up in a media enquiry.
The argument that software must become operable by AI agents, not just usable by people. Pairs directly with the OpenAI agent-ready web standard being tested and the llms.txt v2 spec — see Workflows below.
Matt Wolfe’s daily roundup, leading with the reported Hugging Face sale process and OpenAI’s public backing for a single national AI safety standard over a patchwork of state laws.
OpenAI’s Admin plugin for ChatGPT Work and Codex, read as an enterprise governance development. The issue also explains Anthropic’s invisible text watermarking and what it means for companies reviewing content internally.
A method for using Google’s Gemini Notebook as a comprehension tool — load sources, then interrogate them. Transfers directly to research work in DEVONthink or Obsidian.
ChatGPT for Teachers expands to 55 US school systems, reaching more than 100,000 additional educators and staff with secure tools, training and support.
The research companion to the schools rollout, examining how students and educators use ChatGPT and arguing that support now extends past the boundaries of a class or a term.
Harvard is fronting a paid bootcamp with AI avatars of its own faculty. The issue also reports OpenAI testing a web standard for agent-ready sites, and MIT research finding that chatbot help can weaken a person’s judgment once the chatbot is removed.
A single structured prompt that converts subject expertise into a curriculum, a launch plan and a distribution strategy. Worth stealing for internal enablement material.
Discoveries
Workflows & Tool Watch.
MCP Apps: your connectors can now draw their own interface
The Model Context Protocol — the plumbing that lets Claude talk to DEVONthink, Obsidian, Things 3 and the rest of your stack — published an updated roadmap on 22 August. The headline addition is MCP Apps: a connector can now ship a small interactive web interface that Claude renders inside a secure sandbox, instead of only returning text. In practice that means a DEVONthink or Inoreader connector could hand you a clickable browsing panel rather than a wall of results. Also new: a redesigned “Tasks” extension for long-running work that outlives a single chat, and a Multi Round-Trip Requests pattern that makes back-and-forth prompts work on simpler, stateless servers. The protocol’s official SDKs are now approaching half a billion downloads a month.
RelevantClaude · Claude Code · Cowork · DEVONthink · Obsidian · Inoreader · Things 3
Claude Code’s 26 August release: faster startup and an Auto mode permissions tab
Anthropic shipped a broad stability update to Claude Code. Three things matter for how you actually use it. First, MCP connectors and the sandbox no longer block the first screen from appearing, so sessions with a lot of connectors — which describes yours — start noticeably faster. Second, there is a new Auto mode tab inside /permissions where you can see and edit the rules that decide what Claude is allowed to do without asking. Third, a bug that sent connector tool settings as plain text instead of their real type has been fixed, which was a quiet cause of connectors misbehaving. Background sessions are also smarter and Bash allow-rules now warn you before doing something broad.
Making a website legible to AI agents: llms.txt v2 and the agent-ready web
Two developments this week point the same direction. Jeremy Howard published version 2 of the llms.txt specification — a plain text file at the root of a website that tells AI agents where the important pages are, now with a standard way to serve a clean Markdown version of any page. Separately, Superpower Daily reports OpenAI is testing its own web standard for agent-ready sites. Adoption is already meaningful: 56% of sites surveyed publish a valid llms.txt and 38% publish an AGENTS.md file, but 65% of the top 100 sites still have nothing. For a communications team this is the new equivalent of search optimisation — if a newsroom is not readable by agents, it will be summarised from whatever else is.
Relevantcammacmurchy.com · Tencent newsroom · Media relations · Content distribution
Codex adds session forking, conversation export and cheaper delegation
OpenAI’s coding agent picked up several practical additions: you can now export a conversation from the terminal interface, fork a session and restore it, draft a prompt before startup, and see a credit estimate before running something expensive. It also gained asynchronous hooks that fire alongside MCP connector tools. On price, GPT-5.6 Sol dropped by more than 20% for three months, and cloud task delegation is included on the $20/month Plus plan. GPT-5.4 and GPT-5.4 mini retire from Codex on 31 August — worth knowing if anything you rely on is pinned to them.
An admin control panel for ChatGPT Work — and why comms teams should care
OpenAI’s new Admin plugin lets an organisation see how ChatGPT Work and Codex are actually being used, manage members and permissions, and adjust limits from inside the product. The reason this belongs in a communications briefing rather than an IT one: usage analytics of this kind is the raw material for the “how our people use AI” story that every large company will eventually be asked to tell, and the governance record that regulators will eventually ask to see.
RelevantChatGPT Work · Team coordination · Issues management
Every Claude output now carries a hidden signature — including yours
Worth internalising as a working practice, not just a news item. Anthropic embeds an invisible watermark in text generated by new Claude models, and it applies across Claude, Claude Code, Cowork, Claude Tag and the API, worldwide. The mark survives copy-paste and does not change how the text reads. A detection API is coming. The practical implication for anyone drafting corporate content with Claude: assume any passage that goes out unedited can be identified as machine-generated by a third party. That is an argument for substantive human rewriting rather than light editing, and for being straightforward internally about where AI sits in the drafting process.
RelevantClaude · Cowork · Media relations · Content creation
Gemini Notebook as a reading tool, not a writing tool
The AI-Augmented Engineer’s walkthrough treats Google’s Gemini Notebook as somewhere to load source material and then interrogate it — asking what an argument depends on, where two sources disagree, what is missing — rather than asking for a summary. The same pattern works with a DEVONthink group or an Obsidian folder as the source set, and is a better fit for background research on a media enquiry than a one-shot summary would be.
Published internally on 26 August, the piece reframes the competitive question away from first-mover advantage and toward sustained investment and iteration. It is a useful line to have ready — it answers the “did Tencent move too slowly on AI?” question without conceding the premise, and it is consistent with the capital expenditure trajectory reported at Q2.
As of 26 August, Hy3 has been released and is available globally. Tang Daosheng has said Hy4 will bring more breakthroughs — with the explicit caveat that until it ships it remains an expectation. That caveat is doing deliberate work and should be preserved in any external retelling; stripping it out is how a forward-looking statement becomes a commitment.
Analysys data puts WorkBuddy at approximately 20.97 million visits in Q2 2026, first among the seventeen leading native AI desktop office agents tracked. This is currently the cleanest third-party proof point available for Tencent AI product traction in the office productivity category.
Operating capital expenditure rose 190% year-on-year to RMB 51.8 billion as AI infrastructure investment accelerated, producing negative free cash flow of RMB 13.8 billion. Coverage at the time paired a revenue beat — driven by WeChat advertising and resilient games — with pressure on the shares over AI cost. Expect this pairing (strong top line, questioned AI returns) to be the default frame for the next quarter as well.
Nothing in the last 24 hours of AI coverage names Tencent directly beyond the internal Tang Daosheng piece. The most recent externally reported Tencent AI item remains the 12 August news that a former Alibaba AI architect secured Tencent and HSG backing for a new lab — the line most likely to surface in a reporter’s question about Tencent’s talent and investment strategy in Chinese AI.