Stripe, the payments company, has agreed to acquire OpenRouter — a New York startup founded in 2023 that gives software developers a single doorway to more than 400 different AI models. The point of OpenRouter is cost control: rather than committing to one AI provider, a company can route each task to whichever model does the job cheapest and fastest. The price tag is striking. OpenRouter raised money at a $1.3 billion valuation roughly three months ago; this deal values it above $7 billion, more than five times that figure. The read-across for communications teams is that “which AI model are you using” is quietly becoming an infrastructure and procurement question rather than a branding one — and that the companies building the plumbing between models are now being valued like the model makers themselves.
Anthropic’s chief executive used a lengthy public intervention to concede that public hostility toward AI is a serious problem, and to argue about its cause. Amodei’s diagnosis is that ordinary people do not trust companies, governments or the technology industry, and reflexively assume — in his words — that they are “cooking up some new way to screw them over.” Notably, he pushed back on the idea that the backlash was created by AI executives talking too loudly about catastrophic risk. His prescription is essentially a communications thesis: “less persuasion, more proof.” Trust, he argues, will be earned by AI delivering visible scientific and medical results, not by safety rhetoric or reassurance campaigns. For anyone running corporate communications in this sector, it is a useful framing to have on hand — and a reminder that proof-of-benefit storytelling is outperforming reassurance messaging right now.
Tencent released its Carbon Neutrality Mid-Term Report, setting out progress toward net-zero emissions across operations and supply chain by 2030 — and framing the challenge explicitly against AI’s energy appetite. The report notes that AI workloads are pushing server rack power requirements from a standard 6–8 kW baseline to somewhere between 30 and 100 kW. Two figures are likely to travel: Tencent’s renewable electricity share more than doubled from 22 percent in 2024 to 48.5 percent in 2025, and its next-generation data centre architecture, T-AIDC, is built specifically for high-density AI computing with power-supply efficiency reaching up to 98 percent. The report also describes AI-driven scheduling and digital-twin techniques used to match data centre workloads dynamically to when renewable power is actually available.
OpenAI has begun rolling out Computer History, an opt-in feature in the ChatGPT desktop app for Mac that builds a running timeline of what you do across apps and websites, then converts it into memory ChatGPT can draw on later. Importantly, it does not take screenshots — it records interaction events such as clicks, typing, keyboard shortcuts and app switching, using macOS accessibility APIs, and periodically summarises them into plain-text notes organised by day and time. Activity logs are cached locally for up to 48 hours, then sent to OpenAI’s servers to be turned into high-level text memories, after which the raw activity data is deleted. It is off by default, available to Pro, Business, Work and Enterprise users, and can be paused, app-excluded or wiped. OpenAI says the data is not used to train its models. This is worth understanding before it appears on any corporate Mac fleet — the privacy question is not “is it recording your screen” but “who holds the summary of your working day.”
Google has added a Media Watermark toggle to the Gemini app and its Flow video editor, letting users switch off the visible watermark that has been stamped on AI-generated images, video and music. Josh Woodward, Google’s vice president for Gemini, confirmed the change, which covers the Nano Banana, Omni and Lyria models, with support in Search to follow. Two things stay in place: the invisible SynthID watermark and the C2PA provenance metadata embedded in the file, so it remains technically possible to verify that an asset came from Google’s tools. The timing is pointed — it lands the same fortnight Anthropic began embedding invisible watermarks in Claude’s text output under the EU AI Act. The industry is converging on a split model: visible marks become a user preference, machine-readable provenance becomes the compliance layer.
OpenAI previews Ultrafast mode for GPT-5.6 SolA limited API preview running on Cerebras hardware delivers roughly 750 output tokens per second — up to 14 times standard throughput. Aimed at real-time agents where waiting for the model breaks the experience.
OpenAI · Aug 17
02
Google retires three Imagen 4 model IDsEffective August 17, Google shut down the Imagen 4 generation endpoints and is directing developers to migrate to gemini-3.1-flash-image. Anyone with image generation wired into a workflow should check their model string.
Tencent Hy3 goes global, free through August 31Tencent opened international access to Hy3 via WorkBuddy, Tencent Design Miora and Tencent Cloud TokenHub. It is a 295-billion-parameter mixture-of-experts model that activates only 21 billion parameters per token, shipped under Apache-2.0 with a 256K context window.
Tencent · Aug 2026
05
Perplexity’s Comet browser is now free on every platformPerplexity removed the paywall on Comet, its Chromium-based agentic browser, on July 29. Comet Assistant can browse, fill forms and complete multi-step tasks; Deep Research and external-service agent tasks still sit behind the paid tiers.
Perplexity · Aug 2026
06
Ten new models shipped in August from six providersRelease trackers counted ten model launches this month, including Qwen3.8-27B from Alibaba’s Qwen team on August 14. The pace is now fast enough that model-version references in published material go stale within weeks.
Digital Applied · Aug 2026
Deals & People
07
Anthropic in talks to buy Decart for about $6 billionThis would be Anthropic’s largest known acquisition. Decart, founded in 2023 by Israeli brothers Dean and Orian Leitersdorf with Moshe Shalev, makes software that helps AI chips run more efficiently — cutting the cost of training and operating models. Talks are not finalised and could still fall through.
Bloomberg · Aug 13
08
OpenAI’s IPO: confidential S-1 filed, $1 trillion target, September timelineOpenAI filed confidentially with the SEC on June 8, targeting a listing valued between $852 billion and $1 trillion, with a raise reported at $60 billion or more. Sam Altman has called any valuation below $1 trillion a “nonstarter.” Some analysts think the timeline slips to 2027.
AI Tools Recap · Aug 2026
09
OpenAI appoints Dali Rajic as Chief Revenue OfficerA senior commercial hire as OpenAI builds out enterprise sales ahead of a possible public listing. Worth noting for anyone tracking who now owns the enterprise relationship at OpenAI.
OpenAI · Aug 2026
10
AI funding is at record levels — and heavily concentratedNorth American startup funding broke records in the first half of 2026, driven by AI, but 88 percent of AI capital in Q2 went to US-based companies. Recent large rounds include Hark at $700 million for “personalized intelligence,” HappyRobot at $150 million Series C, and Steno at $49 million Series C.
Crunchbase News · Aug 2026
11
Tencent defends AI spending surge on Q2 earnings callReporting second-quarter results, Tencent flagged a significant rise in capital expenditure and argued its AI investment should generate “superior” returns relative to peers, citing its existing product distribution.
CNBC · Aug 12
Regulation & Policy
12
The EU AI Act’s core obligations are now legally enforcedSince August 2, high-risk system rules and deployer transparency duties are in force, covering risk management, logging and traceability, human oversight and quality management. The EU is the first jurisdiction to impose comprehensive binding AI regulation — and it is now the enforcement era, not the preparation era.
Cubbbix · Aug 2026
13
China fines 12 companies 4.2M RMB in first week of companion-AI enforcementChina’s Implementation Opinions on Intelligent Agents took effect on July 15 and are widely described as the first national framework treating AI agents as a distinct regulatory category. Where the EU’s rules centre on fundamental rights, China’s are built around content control and information security.
AIFOD · Aug 2026
14
OpenAI funds 14 independent AI policy projectsThe company is backing outside research exploring how to expand economic opportunity and strengthen societal resilience as AI spreads. Company-funded independent policy work is a pattern worth watching for how it is received by critics.
OpenAI · Aug 17
15
OpenAI joins the PORTS-Pike project in Southern OhioA community-investment announcement tied to thousands of jobs in Southern Ohio — part of the broader pattern of AI companies pairing data centre expansion with local economic messaging, at a moment when data centres are facing organised local opposition.
OpenAI · Aug 17
Security, Trust & Society
16
OpenAI publishes “The Defender’s Window” on AI and cybersecurityAn argument that AI is reshaping security for attackers and defenders simultaneously, with detail on how OpenAI is hardening its own defences and what security teams should be doing now. Companion to its Daybreak cyber-model work.
OpenAI · Aug 17
17
Researchers find AI-powered hacking tools for sale on underground forumsPurpose-built offensive AI tooling is now being traded commercially in criminal marketplaces — the practical, unglamorous version of the “AI cyber risk” conversation, and the version most likely to reach a corporate security team first.
Cybersecurity Dive · Aug 2026
18
Anthropic’s invisible text watermarks are live — with a detection APISince August 2, Claude subtly biases its word choices to embed a signal detectable across enough text, and signs files with C2PA metadata. It applies worldwide with no opt-out, driven by Article 50 of the EU AI Act. Anthropic has shipped a text-detection API. Heavy editing, paraphrasing or translation can destroy the signal.
Forbes · Aug 13
19
82 percent of journalists now use AI, Muck Rack findsMuck Rack’s State of Journalism report puts ChatGPT use among journalists at 47 percent and Gemini at 22 percent. The same research found 86 percent of journalists say at least some stories originate from PR pitches — but 88 percent immediately discard pitches that miss their beat.
AI Magazine · 2026
20
Nieman Lab reviews a new book on AI rewiring the newsroomA considered look at how AI is changing story ideation, sourcing, verification, storytelling and distribution — and where it is going badly. Useful background reading for anyone whose job depends on how newsrooms actually work now.
Nieman Lab · Aug 2026
From Your Inbox
Substack Highlights.
The dedicated Substack tools were not available in this run, so this section was assembled from the Substack and newsletter feeds arriving through your Inoreader subscriptions. Five newsletters published in the last 24 hours.
Azhar’s Monday roundup of data signals across AI, energy and markets — the format he uses to track the state of the AI economy week by week.
Reports that AI revenues have continued climbing since Exponential View’s June state-of-the-AI-economy research, with the tracking data updated weekly.
Sits alongside his recent pieces on the AI capex logjam and on the market misreading Google’s DeepMind departures — a consistent argument that headline spending numbers are being read too simply.
Opens on Amazon’s reportedly enormous AI spending and works through what it costs to actually run an agent at scale.
Exponential View is recruiting an AI Economy Research Fellow — a signal of how much of the publication’s output is now original economic research rather than commentary.
The through-line: agent pricing is collapsing faster than the infrastructure bills behind it, and someone eventually absorbs the gap.
Leads on AI voice-cloning fraud being used against retail banking customers — the consumer-facing edge of the synthetic media problem.
Also covers how Anthropic’s Claude watermarks actually work, and Google’s decision to let users switch its visible watermark off.
See also · Top Stories
From Your AI Feeds
Inoreader AI Folder.
Eleven articles landed in your “AI” folder in the last 24 hours. The Stripe–OpenRouter deal and Dario Amodei’s trust remarks dominated the newsletter feeds; OpenAI published three separate corporate posts in a single day; and the GDELT Project ran two pieces on using Gemini for research work that are worth your time.
The GDELT Project already publishes daily Media Trends reports analysing themes and narratives across global television news. Leetaru describes bolting Gemini Deep Research onto that pipeline to automatically generate contextual backgrounders, so a journalist or geopolitical researcher reading the trend data also gets the history behind it. This is the clearest practical template in today’s feed for turning a monitoring feed into a briefing product.
A step-by-step account of using Gemini to work backward from a 1986 reference book consulted in the Library of Congress reading room to locating the actual source images in digitised collections. A good demonstration of AI as a research navigation tool rather than a content generator — the model does the finding, the human does the judging.
A build tutorial for a desktop application that turns meeting conversations into automatic CRM updates and then triggers downstream workflows from them. The wealth-management framing is incidental; the underlying pattern — meeting audio in, structured relationship records out — maps directly onto how a personal CRM like Mesh could be fed automatically rather than by hand.
OpenAI’s position paper on how AI is reshaping cybersecurity for attackers and defenders at the same time, including what the company is doing to strengthen its own defences and what it recommends security teams do now.
The day’s biggest deal news arrived through two of your newsletter feeds simultaneously, both leading with the same story and pairing it with reporting that Anthropic may be holding back an unreleased model.
Coverage of Amodei’s argument that AI’s public-perception problem is a trust problem, plus the Decart bid and a report on an AI hacking tool being sold through Amazon.
Azhar’s weekly data roundup across AI, energy and markets, including updated tracking on AI revenue growth since Exponential View’s June research report.
Voice-cloning fraud aimed at banking customers, alongside explainers on how Claude’s watermarks work and Google’s new watermark toggle.
See also · Top Stories · Substack Highlights
Discoveries
Workflows & Tool Watch.
Turn a monitoring feed into an auto-briefed intelligence product
Kalev Leetaru at the GDELT Project has wired Gemini Deep Research into GDELT’s daily Media Trends reports so that every trend analysis automatically ships with a contextual backgrounder explaining the history behind the narrative. This is the single most transferable idea in today’s briefing for a global comms function. You already run media monitoring; the missing piece is usually context for whoever reads it at 7am. The pattern is: monitoring output goes in, a deep-research pass generates the “why this matters and what came before” section, and the combined document goes out. It would sit naturally on top of your existing Inoreader and DEVONthink pipeline.
The MCP standard just got a major rewrite — and the connector directory passed 950 servers
MCP, or Model Context Protocol, is the plumbing that lets an AI assistant talk to your other software — it is why Claude can reach into Inoreader, DEVONthink, Obsidian or Things on your behalf. The 2026-07-28 specification is now final, and it changes the architecture in a way that matters practically: MCP moved from a stateful, always-connected protocol to a simple request-and-response model, which means MCP servers can now run on serverless and edge infrastructure rather than needing something running constantly. It also brings stronger OAuth and OIDC sign-in and versioned extensions for Apps and Tasks. In plain terms: expect many more hosted connectors that you do not have to run yourself. Anthropic’s directory now lists more than 950 servers.
RelevantClaude · Claude Code · Cowork · DEVONthink · Obsidian · Things 3
Meeting audio in, CRM records out — a build pattern worth stealing for Mesh
Recall.ai published a walkthrough for building a desktop application that listens to meetings, extracts what matters, and writes it straight into a CRM as structured records, then triggers follow-on workflows from those updates. The tutorial is framed around wealth management, but the mechanism is general. Given how much of your week runs through VooV, Zoom and Teams, and given that Mesh only becomes valuable when it is actually current, this is the shape of automation that closes the gap — you would pair it with MacWhisper or PLAUD for capture rather than building transcription from scratch.
RelevantMesh · MacWhisper · PLAUD · Zoom · VooV · Microsoft Teams
Anthropic’s watermark detection API is a practical tool for a comms team
Alongside embedding invisible watermarks in everything Claude writes, Anthropic has shipped an API that lets you check whether a given piece of text carries that signal. For a communications function this is genuinely useful — vetting agency copy, checking submissions, or verifying material before it goes into a release. Two caveats you should know before relying on it: the signal degrades or disappears if the text has been heavily edited, paraphrased or translated, and short passages may simply not contain enough words to give a reliable reading. It confirms Claude authorship; it cannot prove human authorship.
n8n’s 2.0 line added two nodes that change what is possible in a self-hosted automation setup: an MCP Server Trigger and an MCP Client Tool. The first lets an n8n workflow present itself to Claude or another assistant as a tool it can call; the second lets an n8n workflow reach out and use tools on remote MCP servers. Practically, that means a workflow you have already built in n8n becomes something you can simply ask Claude to run. Zapier is coming at the same problem from the other direction with Zapier Agents running across 8,000-plus apps and a Copilot that builds automations from a plain-English description.
Claude Code’s August changelog: GitLab merge requests, safer Linux builds
This month’s Claude Code updates add GitLab merge request support to the worktree flag and the agents view, faster self-hosted runner starts, stronger gateway and plugin validation, and improved accessibility and diagnostics. The quietly useful one is opt-in memory cgroup support for Bash commands on Linux, which stops a runaway build from freezing an entire session. Not directly your daily workflow, but relevant if anyone on your team is running Claude Code against internal repositories.
Perplexity Comet is free everywhere — and it runs Chrome extensions
Perplexity dropped the paywall on Comet, its agentic browser, at the end of July. Two details make it worth an experiment: it is built on Chromium, so your existing Chrome extensions and 1Password integration carry over intact, and Comet Assistant can complete multi-step tasks — browsing, filling forms, working through a process — rather than just answering questions. Deep Research and agent tasks that touch external services still require a paid tier. Given how much of your research day happens in a browser, this is a low-cost thing to run alongside your current setup for a week.
RelevantPerplexity · 1Password · Research workflow
Computer History as a personal-context engine — with a governance question attached
Setting the privacy debate aside for a moment, OpenAI’s Computer History is an attempt to solve a real irritation: having to re-explain your context to an assistant every single session. It builds a plain-text timeline of your working day from interaction events, not screenshots, and feeds that back as memory. If you try it, the settings that matter are app exclusions, the pause control, and deletion. The governance question for a corporate environment is not whether it is technically safe but whether a summarised record of an executive’s working day should sit on a third-party server at all — worth a conversation with IT security before it appears on managed devices.
RelevantChatGPT desktop · macOS · Corporate IT policy
The report is the day’s most prominent Tencent story and it is being picked up specifically through an AI-energy lens. The reusable facts: renewable electricity share rose from 22 percent in 2024 to 48.5 percent in 2025; the T-AIDC data centre architecture is purpose-built for high-density AI computing and reaches up to 98 percent power-supply efficiency; AI-driven scheduling and digital twins are used to align compute workloads with renewable availability, and in steel production to optimise energy use. The framing device coverage keeps reaching for is that AI is pushing rack power from a 6–8 kW baseline to 30–100 kW — expect that number in any follow-up question.
Following Hy3’s release on 6 July, international access is now open via WorkBuddy — free to users and enterprises worldwide until 31 August 2026, Pacific Time — as well as through Tencent Design Miora and Tencent Cloud TokenHub. Technical specifics likely to come up: 295 billion parameters total with only 21 billion activated per token, Apache-2.0 licence, 256K context window. The positioning line in the materials is “practical AI” — stable performance across software development, office productivity, financial modelling, front-end design and game production, with reduced token consumption per task. This is also the first major model release since Tencent recruited Yao Shunyu from OpenAI.
Reporting from the second-quarter results is still circulating, with CNBC’s framing — a spending surge defended on the basis of potentially superior AI returns — the version most likely to be quoted back at you. The underlying argument, that Tencent’s existing product distribution lets it convert AI investment into usage faster than peers, pairs naturally with the Hy3 free-access window and the T-AIDC efficiency numbers if you need a single coherent narrative across all three.
Nothing negative surfaced in the monitored feeds today. The two live threads a reporter is most likely to pull on are the energy cost of AI expansion, which the Carbon Neutrality report is designed to answer directly, and the capex-versus-returns question from the Q2 call. Both have published company positions to point to.