Anthropic released Claude Opus 4.8 today, arriving less than two months after Opus 4.7 and ramping up the company’s upgrade cadence. Anthropic describes the model as having “sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors.” On SWE-Bench Pro, a benchmark for coding agents, Opus 4.8 scored a record 69.2 percent versus 4.7’s 64.3 percent. Pricing is unchanged, and Fast mode is now three times cheaper than on previous models.
Cognition, maker of the autonomous coding agent Devin, closed a $1 billion round at a $26 billion post-money valuation late Wednesday — a more than 2.5x jump from its September 2025 mark. Annualised revenue has climbed from $37 million to $492 million in 12 months, and enterprise usage has grown 50 percent month-on-month for six straight months. Lux Capital, General Catalyst and 8VC led the round, with Mercedes-Benz, NASA, Goldman Sachs and Santander already on the customer list. Cognition says 90 percent of its own code is now written by Devin.
Robinhood launched beta support for AI agentic trading on Wednesday, allowing users to create separate sub-accounts that an AI agent can use to analyse a portfolio, suggest trades and execute orders against a pre-loaded balance. The company also unveiled a virtual credit card scoped specifically for AI agents, with user-set monthly limits and optional per-purchase approval. Stocks come first; options, crypto, futures and prediction markets are flagged as next. American Banker called the move a “wake-up call” for banks.
China is requiring government approval for overseas travel by select founders, researchers and executives working on advanced AI at private companies including Alibaba and DeepSeek, according to Bloomberg. The policy was first applied quietly to some DeepSeek executives in December 2025 and has been broadened in recent months. Authorities reportedly target individuals by strategic value rather than seniority. The expansion signals how seriously Beijing now treats private-sector AI as a national asset — a development worth tracking for anyone in China comms.
Pragmatic Engineer’s Gergely Orosz coined “tokenmaxxing” to describe employers rewarding staff for maximum AI usage measured in tokens. New reporting this week, amplified by Gary Marcus’s Substack, shows the trend is reversing as budgets tighten — Uber’s CTO confirmed the company burned its entire 2026 AI budget in four months, and ServiceNow blew through its full-year Anthropic budget in months. Under “best case” assumptions, Microsoft AI ROI is at -9 percent, Google -15 percent, Meta -28 percent and Oracle -35 percent. Anthropic still leads OpenAI on revenue, but the runway may be shorter than headline numbers suggest.
Pivot To AI: “The SpaceX IPO works like a crypto fraud, but with AI”Critical commentary from David Gerard and Amy Castor arguing the structure of the upcoming SpaceX IPO echoes the playbook of crypto initial coin offerings — promise machinery, insider rug-pull, retail exit liquidity.
Pivot To AI · May 28
Regulation & Governance
13
OpenAI publishes its “Frontier Governance Framework”A new document laying out how OpenAI’s safety, security and risk practices align with emerging EU and California regulations. A useful primer for any comms team that fields regulatory-alignment questions.
A dedicated Substack tool isn’t connected today; these highlights are drawn from the Substack newsletters Inoreader pulled in via RSS over the last 24 hours, supplemented by direct web reads of the publications themselves.
“Tokenmaxxing” — employees using AI as much as possible to hit internal targets — has been padding Anthropic and OpenAI’s revenue, but Marcus says the trend is rolling over.
He cites best-case AI ROI of -9 percent at Microsoft, -15 percent at Google, -28 percent at Meta and -35 percent at Oracle, with only Amazon barely positive.
Marcus’s bottom line: “Brace for bailouts” — particularly worth tracking for anyone briefing executives on the AI investment cycle.
Critiques the assumptions underpinning Anthropic’s recent research on US-China AI competition timelines.
Argues Western frontier-lab arguments often treat Chinese AI ecosystems as monolithic and ideologically rigid, ignoring the diversity inside Alibaba, Tencent, ByteDance and DeepSeek.
Useful counter-positioning for any Tencent-related media brief that touches the “US vs China AI race” frame.
swyx lays out Latent.Space’s 2026 plan: AINews daily, the weekly LS Pod, and broader LS Essays consolidated on Substack.
Announces Latent.Space is becoming a podcast network, starting with a new “AI for Science” show.
Notes AI Engineer conference is scaling from 4 events in 2025 to at least 7 in 2026 — a useful sense of where the practitioner community is congregating.
Tesla disclosed plans for a dedicated Optimus humanoid robot factory — investor commentary in the issue centres on whether unit economics work below $20K per robot.
Sidebar coverage of China’s overseas travel restrictions on AI researchers, with the AI Valley take that the curbs will accelerate the exodus of Chinese AI talent to Singapore.
Provocative read on a circulating thesis that several big-tech CEOs are increasingly believing their own AGI narratives.
Plus: agentic AI and AI inference now drive 450 percent more web traffic than a year ago, putting CDN and bot-management costs squarely on comms teams’ radar.
Procurement-angle read on the Cognition round: argues the $492M run-rate is the number that actually matters to enterprise procurement teams, not the $26B valuation.
Covers five other moves this week including Robinhood and Microsoft Copilot Cowork.
See also · Top Stories (Cognition).
From Your AI Feeds
Inoreader AI Folder.
14 articles in your “AI” folder in the past 24 hours. Today’s mix leans toward enterprise agent launches, China AI talent geopolitics, and end-of-week newsletter wrap-ups.
David Gerard and Amy Castor argue the SpaceX listing is structured to mirror the crypto ICO playbook — impossible white-paper claims, price spike, insider exit — and now uses AI hype as the recurring “number-go-up” rationalisation.
OpenAI’s official case study with consultancy Endava: Codex compressed requirements analysis from weeks to hours, with quoted leadership describing internal restructuring around agent-supervisor roles.
Kalev Leetaru revisits an experiment from a year ago that used Gemini 2.5 to analyse two 500-year-old choir book leaves — a side-by-side with current Gemini models to measure how much vision-and-historical-context capability has improved in 12 months.
Companion post applying Gemini to a small Asiatic leaf collection from a History of Writing & Printing subcollection — useful illustration of where multimodal models help with cultural heritage and archival workflows.
The Neuron is running a live “Agents for total beginners” session today aimed at non-technical knowledge workers — relevant for any comms or PR team thinking about agent literacy training.
Tesla announces a dedicated Optimus factory; the issue’s secondary item is China’s overseas travel restrictions on AI researchers at Alibaba and DeepSeek.
Daily AI news roundup covering Cognition’s billion-dollar round, Robinhood agents, OpenAI/Endava, and reflections on the proliferation of AI-wrapper SaaS products.
Practical paid-subscriber guide to the redesigned Claude Code desktop app — flag-by-flag walkthrough of the new dispatched-session options and how to keep token usage from running away.
“A coding-agent company posted a $492 million revenue run-rate alongside its $1 billion round, and the run-rate is the number that matters for procurement teams.” Plus five other things that moved.
OpenAI publishes its full governance framework explaining how its AI safety, security and risk practices align with emerging EU and California regulations.
Read on the “AI psychosis” thesis circulating in tech press — that founders increasingly believe their own AGI narratives. Plus data showing agentic AI now drives 450 percent more web traffic than standard AI.
The Neuron’s flagship daily covering Robinhood agentic trading, AxiomProver, Google’s new cyber agents, MagicPath, and AI music tools.
See also · Top Stories · Substack Highlights
Discoveries
Workflows & Tool Watch.
Claude Code desktop now runs “tens to hundreds” of background agents in parallel
The redesigned Claude Code desktop app shipped Wednesday lets you describe an outcome and have Claude orchestrate parallel agents in the background while you keep working. The /usage breakdown now shows large session files (helpful for catching expensive ones), thinking summaries stay readable for at least three seconds, and macOS background agents now show up properly in Privacy & Security so permission grants survive upgrades. Practical use case: a long-running comms research workflow that fans out across web search, document analysis and draft generation, with the agents reporting in.
Obsidian 1.12.7 ships a CLI that Claude Code can drive directly
The latest Obsidian release exposes its internal commands (file creation, link management, settings updates) as terminal-callable operations. The practical upside: Claude Code, Cowork mode and any MCP-aware tool can drive Obsidian as a first-class app instead of just shoving Markdown into a folder. For a vault-based note-taker, this is the cleanest path yet to have an AI assistant actually maintain your second brain — daily summaries, meeting notes, project rollups — bidirectionally.
Perplexity Comet enterprise: silent MDM deployment plus model swap behind the browser agent
Comet for Enterprise lets IT admins deploy the browser silently across macOS and Windows via MDM and configure hundreds of policies. The new model picker lets Max subscribers swap Opus 4.6 or Sonnet 4.5 behind the Comet Browser Agent for tasks like analysing a dashboard, walking through a competitor’s onboarding flow or pulling structured data from a GitHub commit history. If you’ve been waiting to roll out a browser-level AI agent to a comms team, the rough edges are off.
RelevantPerplexity Comet · macOS · Claude Opus 4.6
n8n’s MCP server lets Claude build and iterate on workflows from your terminal
n8n’s official MCP server now lets Claude Code build, test and refine n8n automation workflows without leaving the terminal. Combined with n8n’s native AI Agent node (which can use any other n8n node as a “tool”) this means you can describe a workflow in plain English — “every morning at 6am pull yesterday’s Inoreader items in the AI folder, summarise them with Claude, and post a digest to a Notion page” — and have Claude scaffold and validate the n8n scenario for you.
Microsoft Copilot Cowork: autonomous multi-step agent across Word, Excel, PowerPoint and Outlook — built on Claude
Microsoft and Anthropic shipped “Copilot Cowork” — a new agent inside Microsoft 365 that executes complex workflows across the Office apps without prompting at each step, powered by Claude under the hood. For a global comms team that lives in Outlook, PowerPoint and Excel, this is the most credible Office-native agentic experience yet. Worth a pilot before quarterly reviews.
RelevantMicrosoft 365 · Claude · Outlook · PowerPoint · Excel
StackAdapt’s MCP server brings campaign intelligence into Claude
StackAdapt launched an MCP server that exposes campaign performance, audience and creative data directly to Claude. For a comms or PR team that runs paid amplification alongside earned coverage, having Claude pull StackAdapt numbers next to Inoreader clip counts and Slack signal in a single conversation is a meaningful workflow simplification.
“Persistent context + bidirectional memory” — the AI vault pattern getting traction in 2026
A workflow gaining adoption this year: keep your AI assistant grounded in your Obsidian or DEVONthink vault as persistent context, write decisions and corrections back into the vault as the conversation runs, and point new “skills” at vault reference files instead of stuffing context into every prompt. The pattern is particularly useful for a long-running comms job where context (talking points, prior statements, executive preferences) accumulates over years.
No fresh, direct Tencent breaking news in the past 24 hours, but the FY-2026 AI spend pledge (more than doubling 2025’s RMB 18 billion) remains the most-cited Tencent line in industry coverage. President Martin Lau set that number out at Q1 results on May 13. Expect the next round of reporter questions to focus on what proportion is going into Hunyuan Hy3 versus Yuanbao versus WorkBuddy, and how the figure compares with ByteDance’s $23 billion AI infrastructure plan and Alibaba’s RMB 123 billion 2025 capex.
Background line to have ready: Hy3 is a 295B-parameter mixture-of-experts model with 21B active parameters, and has held the top spot in its parameter class on OpenRouter token measurements since late April. WorkBuddy continues to be positioned as the most widely used productivity AI agent in China.
Jeffrey Ding’s argument that Western frontier-lab research often treats Chinese AI ecosystems as monolithic gives a useful frame for any Tencent comms response if a journalist tries to lump Tencent’s roadmap into a generic “China AI race” narrative.