AI Intelligence Report
Top Stories
ChatGPT moves into personal finance — and into your bank accounts
OpenAI launched a preview of personal finance tools inside ChatGPT for Pro subscribers in the United States yesterday, partnering with Plaid to let users connect more than 12,000 financial institutions including Schwab, Fidelity, Chase, Robinhood, American Express and Capital One. Once connected, ChatGPT shows a dashboard of portfolio performance, spending, subscriptions and upcoming payments, and answers natural-language questions about budgeting and future planning. The launch comes a month after OpenAI acquired the team behind the personal-finance startup Hiro, and OpenAI says more than 200 million users already ask financial questions of ChatGPT every month. For communications professionals, this is a significant signal: a general-purpose AI is now competing directly with banks, brokerages and personal-finance apps for the first conversation a customer has about their money. Bloomberg reports the company is also weighing additional capital raises to meet AI demand.
PwC and Anthropic expand alliance — Claude Code and Cowork going to hundreds of thousands
Anthropic and PwC announced an expansion of their strategic alliance on May 14, with PwC rolling out Claude Code and Cowork mode to its U.S. teams first, then expanding toward a global workforce of hundreds of thousands of professionals. Anthropic says Claude is already in production at PwC across professional sports operations, insurance underwriting, mainframe modernization, HR transformation and cybersecurity, with delivery times cut by up to 70 percent in some workflows. This is the second major Anthropic enterprise expansion this week — the company also committed $200 million in grants and credits to a four-year partnership with the Gates Foundation on global health, life sciences, education and economic mobility.
Tencent’s Q1 results put AI front and centre — Hy3 leads, AI spend more than doubling
Tencent reported Q1 2026 results on Tuesday with several significant AI updates flagged for investors. The Hy3 preview model, built by Tencent’s revamped AI research team on re-architected infrastructure, was described as a leader in its parameter-size class and has been top-ranked in OpenRouter token measurements since April 28. The company said its productivity AI agent solution WorkBuddy is currently the most widely used productivity AI agent service in China, and that its automated campaign-management solution AIM+ powered approximately 30 percent of total advertiser marketing-services spending. Tencent spent RMB 18 billion on its Hunyuan foundation model and Yuanbao AI assistant in 2025 and committed to more than doubling that to in excess of RMB 36 billion in 2026.
Google’s “Googlebook” and the Gemini-everywhere push ahead of I/O
Ahead of Google I/O 2026, Google is pulling out all the stops, with Gemini now deeply integrated into Chrome for contextual summarisation, smart form-filling and real-time translation across any webpage. Google also unveiled Googlebook, an AI-native laptop platform pitched as the long-term successor to Chromebooks, designed around Gemini and featuring a “Magic Pointer,” AI-generated widgets and proactive task assistance. Android is being repositioned to act more like an agent — Gemini will pull relevant information from Gmail, build shopping carts and book reservations across apps. A leak of Gemini’s “Spark” agent surfaced in newsletter coverage yesterday, suggesting more agent capability is queued for I/O.
OpenAI reportedly preparing legal action against Apple; iOS 27 to open up to third-party AI
On Thursday, Bloomberg reported that OpenAI has retained outside counsel and is considering legal options including a formal breach-of-contract notice against Apple over a ChatGPT integration that has not delivered the subscriber numbers or prominence OpenAI expected. The reporting lands as Apple is preparing a major platform shift in iOS 27, iPadOS 27 and macOS 27 — internally called “Extensions” — that would let users pick third-party AI providers such as Google and Anthropic to power Apple Intelligence features. For a global communications team, the takeaway is that the most consequential AI integration story of the next 12 months may not be a model release but a venue fight over who gets to be the default assistant on a billion-plus devices.
EU AI Omnibus deal: high-risk AI rules postponed, AI Act simplified
At 4:30 a.m. on May 7, EU legislators agreed to the AI Omnibus — the regulation that amends the EU AI Act — closing six months of tense negotiation. The headline change is a delay in the application date of requirements for high-risk AI systems, giving developers and regulators more time to set up compliance machinery. On the U.S. side, Connecticut’s bipartisan SB5 — a broad omnibus AI package covering employment, chatbot safety, synthetic content labelling and anti-discrimination — passed both chambers on May 1 and is expected to be signed by Governor Lamont. Fortune flagged on Thursday that the U.S. now has roughly 1,200 active AI bills and no agreed framework for measuring them.
AI News Roundup
Substack Highlights
Inoreader AI Folder
arXiv bans academic authors for AI slop papers
David Gerard reports that arXiv, the preprint server academics rely on to publish results quickly, has begun banning authors who flood the system with low-quality, chatbot-generated papers — the so-called “AI slop” problem. Gerard’s frame: the chatbots have escalated a manageable spam problem into a structural one, threatening the trust scientists place in the platform.
A new personal finance experience in ChatGPT
OpenAI’s own announcement of the ChatGPT–Plaid personal-finance preview for Pro users — connect your accounts and ChatGPT will analyse your spending and portfolio.
Databricks brings GPT-5.5 to enterprise agent workflows
Databricks uses GPT-5.5 for enterprise agent workflows after the model set a new state of the art on the OfficeQA Pro benchmark. Notable because Databricks customers are large regulated enterprises, so this is a useful proxy for where frontier-model adoption sits in the Fortune 1000.
The data center in your house
Matt from Future Tools covers the wave of consumer “AI inference appliances” coming to market and Amazon’s plan to replace its Rufus shopping assistant with a more capable agent. Frames the home-AI-server pitch against the existing OpenAI/Google/Anthropic cloud incumbents.
Gemini’s Spark agent has leaked
Leak coverage of Gemini’s “Spark” — Google’s next agent layer for Android, expected to be formally announced at I/O. Same letter flags the OpenAI–Apple legal-action reporting as a developing story.
Codex is now something you can manage from your phone
AI Breakfast covers OpenAI moving Codex into the ChatGPT mobile app. The practical implication: long-running coding tasks become asynchronous, phone-monitored work — the same pattern Anthropic shipped earlier with Claude Code on mobile.
Mintlify just killed seat pricing in one sentence
The newsletter walks through three same-week moves from SaaS companies (Mintlify is the centerpiece) abandoning per-seat pricing because “agents are the users.” Argues the entire SaaS pricing model is being rewritten — relevant for anyone forecasting enterprise-software budgets in 2026–27.
The Prompt That Helped Me Build My First Digital Product
The author shares a single structured prompt that generated a product idea, positioning and a launch plan in under ten minutes. The template is reusable for press-pitch ideation and campaign-concept work.
Americans hate AI data centers more than nuclear plants
Polling shows U.S. public opinion has turned harder against AI data-centre construction than against new nuclear plants. The piece sits alongside an OpenAI security-breach note and is being read as evidence that data-centre siting will become a political flashpoint in 2026 midterms.
Lore Issue #184: Isomorphic Raises $2.1B to Solve Disease With AI
Lore’s Friday letter covers the Isomorphic Labs raise, plus shorter items on Google rethinking the mouse pointer (the “Magic Pointer” we saw in the Googlebook coverage), Krea launching its first foundation model, and Figure robots working full shifts.
OpenAI put Codex in the GPT mobile app
The Neuron’s daily letter leads with Codex on mobile, and flags reported U.S.–China AI safety talks as a parallel storyline — framed as a possible thaw in the AI cold-war narrative.
Cerebras and the IPO pop
Azeem Azhar’s Friday note on the Cerebras listing — see Substack Highlights above for the longer summary.
Kevin O’Leary’s giant Utah data centre: will it happen?
Gerard reports the county commissioners have approved O’Leary’s “Stratos Project,” a 9-GW AI data centre on land he’s been quietly acquiring — with vocal opposition from local residents. Pairs neatly with the polling data in The Automated.
Work with Codex from anywhere
OpenAI’s own product post on Codex mobile — monitor, steer and approve coding tasks in real time across devices and remote environments.
AI Workflows & Tool Watch
Claude Code’s May 14 update — Opus 4.7 by default, MCP fixes, plugin SKILL.md improvements
Anthropic shipped a notable Claude Code update Thursday. Fast mode now uses Opus 4.7 by default (up from Opus 4.6) — meaning your everyday Claude Code sessions are quietly running on a smarter model. The update also fixed two MCP-server bugs that have been silently breaking longer-running tool calls: the per-request MCP fetch timeout was capped at 60 seconds regardless of your MCP_TOOL_TIMEOUT setting (now fixed), and stdio MCP servers were receiving corrupted arguments when shell prefixes contained spaces. A small but useful change: plugins with a root-level SKILL.md and no skills/ sub-directory are now picked up automatically. Plain-English version: Claude Code talks to outside tools through small connectors called “MCP servers” — these fixes mean those connections are more reliable for tasks that take longer than a minute, like processing a long document or hitting a slow internal API.
Spark Mail adds a Mac CLI and “agent skills” for Claude Code, Codex and others
Readdle quietly shipped a major update to Spark Mail two weeks ago that lets coding agents like Claude Code and Codex read your email, calendar, contacts and meeting notes through a Mac command-line tool. MacStories has the practical walkthrough. Why it matters for a comms team: this is the first mainstream email client where you can write something like “draft replies to anything from a reporter in the last week, prioritise by deadline” and have the agent read the inbox itself. Read-only by default, which sidesteps the obvious safety concern.
Perplexity Comet on Mac, with Opus 4.6 as the browser-agent default
Two Perplexity updates worth noting in the past week. On May 4, Perplexity made Opus 4.6 the default model behind its browser agent for Max subscribers (Sonnet 4.5 is also available) — they say it noticeably improves complex tasks like reviewing a competitor’s onboarding flow or analysing GitHub commit history. On May 7, Perplexity Computer — their full-fat agentic Mac app that can drive your browser and act on tasks — became available to all Pro Mac users, not just Max. Enterprise admins can deploy it silently across macOS and Windows via MDM. For Cam: the Mac app is the most interesting demonstration to date of an AI that genuinely uses your computer the way you do, not just a chat box.
Codex shortcuts library lands for iOS Shortcuts
Matthew Cassinelli published a new shortcuts library yesterday for OpenAI’s Codex, including a “Create Task” action that lets you fire Codex jobs directly from any iOS automation. Useful for stringing together: “transcribe this voice memo with MacWhisper → run a Codex task to clean it up → save to Drafts.”
Automation Academy: “How I Can Automate Anything I Want Using Codex for Mac”
Federico Viticci’s MacStories Club piece walks through using Codex on macOS as a general-purpose automation runtime — running Apple Script, hitting REST APIs, generating Shortcuts on the fly. Behind a Club paywall but worth it if you’re rebuilding the team’s automation stack this quarter.
Obsidian + n8n + Claude: local agentic note-sorting overnight
A new Obsidian Forum showcase walks through a local agentic system that sorts your fleeting notes while you sleep — built with Obsidian, n8n and Claude. The pattern: hit a button before bed, wake up to processed, tagged and routed notes. The same plumbing works for clipping inboxes, briefing folders or any “I’ll deal with this later” pile.
A dictation app with a CLI — Monologue for iOS/Mac
Viticci flagged Monologue, a new dictation app with a real command-line interface — meaning you can pipe its transcript output into other tools without copy-paste. Pairs naturally with MacWhisper for one-off transcription and gives a clean “voice memo → cleaned text → routed to draft” pipeline.
GitHub MCP Server adds secret scanning at GA
GitHub announced the general availability of secret-scanning support through its MCP server — so any AI tool that speaks MCP (including Claude Code and Cursor) can now ask GitHub directly about exposed credentials. Plain-English version: MCP is a standard way for AI tools to plug into outside services, like a universal adapter. This is the kind of “boring infrastructure” that quietly makes AI safer to use inside a large organisation.
Tencent Mentions
Tencent Q1 2026 results — strong AI commentary, RMB 36B+ AI spend for 2026
Filed Tuesday. The most flagged items for AI coverage: Hy3 preview model leading its parameter class and ranked at the top of OpenRouter token usage since April 28; WorkBuddy positioned as the most-used productivity AI agent in China; AIM+ driving roughly 30% of marketing-services advertiser spend; and the headline 2x increase in foundation-model investment to north of RMB 36 billion.
Analyst framing: AI and core games are the drivers
TipRanks’ read of the Q1 numbers — AI and the gaming portfolio are the two engines analysts are calling out. Useful to have handy if a reporter wants the consensus view.
Caixin: Hy3 framed as the first model from “Tencent’s AI rebuild”
Caixin’s framing — Hy3 is the first major model output of Tencent’s restructured AI team and infrastructure. Worth tracking for how that “rebuild” narrative is settling in Chinese-language coverage versus the international press.
Tencent Cloud AI Coding Challenge wrapped in Singapore
The Singapore station of the 2026 AI Coding Challenge concluded May 9. Solid “AI for Good” angle in regional coverage — worth pinning if the team wants regional case studies on hand.
Tencent on scenario-based AI rollout for industry
Tencent’s own post on globally rolling out scenario-based AI capabilities aimed at industrial-efficiency use cases — the kind of B2B framing that pairs well with the Q1 numbers.
Indirect Tencent watch
James Fallows’ “Trump Lands in Beijing” piece (Sinica, May 13) and Caixin’s recurring coverage of U.S. export-control conversations are the two threads most likely to put Tencent’s name in U.S. headlines next week. Worth watching closely.
Generated by Claude · AI Intelligence Report for Cam MacMurchy
Sources include web search, 143 Substack subscriptions, Inoreader AI folder, Reddit, Discourse forums, Hacker News, and Product Hunt.