AI Intelligence Report
Top Stories
Pentagon strikes AI deals with seven Big Tech companies after shunning Anthropic
The U.S. Department of Defense awarded major AI contracts to seven leading technology firms over the past 72 hours, but pointedly excluded Anthropic — the company most associated with safety-first AI — from the round. Anthropic’s CEO Dario Amodei was reportedly invited back to the White House late last week as the administration seeks to repair the relationship. The move is being read in Washington as a signal that the Pentagon is prioritising raw capability and existing vendor relationships over the careful-deployment posture Anthropic has built its brand on.
OpenAI rolls out GPT-5.5 just six weeks after GPT-5.4, accelerating frontier race
OpenAI’s newest model, GPT-5.5, is now widely available and is reportedly stronger at coding, “computer use” (controlling a browser or desktop on your behalf), and long-running research tasks. In a controlled cyber-security test, GPT-5.5 matched Anthropic’s “Mythos” tool, solving a 12-hour expert-level challenge in under 11 minutes for $1.73 in compute — a benchmark that has alarmed cybersecurity leaders. Separately, OpenAI loosened its long-standing exclusivity arrangement with Microsoft last week, opening the door for distribution through Amazon and possible ties with Google.
Anthropic takes Claude Cowork out of preview and puts it in front of the enterprise
Anthropic graduated Claude Cowork — its desktop “agent for non-developers” mode — out of research preview earlier this month, and is now actively pitching it to large enterprises. New connectors shipped in the last day or two cover Google Drive, Calendar, Gmail, DocuSign, Apollo, Outreach, WordPress, FactSet, Clay and Harvey, among others. For comms teams, the meaningful change is that Claude can now reach into the tools people actually live in (email, Drive, Slack, content systems) without engineering work. Read the New Stack write-up.
Tencent pledges higher AI investment in 2026 as compute scarcity drives Cloud price hikes
Tencent confirmed plans to step up AI capex this year despite chip-export curbs, and Tencent Cloud announced a third price increase in 2026 — its CodeBuddy AI coding assistant goes up 154%, from 78 yuan to 198 yuan per user per month, effective May 15. Multiple analysts say the industry has shifted from a “discount war” to a “price hike cycle” because inference demand is outstripping hardware supply. The same day, Tencent Cloud rolled out enterprise-grade Agent products including ClawPro Private Cloud, ADP Intelligent Workbench, Agent Memory and Agent Storage. Flagged for company watch.
HSBC Research: Tencent’s Hunyuan 3 preview narrows the gap with top Chinese models on coding and agents
HSBC published a research note over the weekend saying Tencent’s Hunyuan 3 (HY3) preview release “trims the gap” with the leading Chinese frontier models on coding and agentic capability — a notable inflection given how far behind Hunyuan was perceived to be at the start of 2026. The note also flags Tencent and Alibaba as “direct beneficiaries” of DeepSeek V4 reaching open-source state-of-the-art, since both companies host and serve those models commercially. Cross-reference: the open-source DeepSeek-V4-Pro now beats every rival open model on math and coding and only trails Google Gemini 3.1-Pro on world knowledge.
Google commits up to $40B more to Anthropic; Meta launches “Muse Spark” from Superintelligence Labs
Google announced an additional Anthropic investment of up to $40 billion — $10 billion immediately, $30 billion tied to milestones — alongside expanded long-term compute support. Meanwhile, Meta unveiled Muse Spark (formerly code-named Avocado), the first model out of Meta Superintelligence Labs since the Alexandr Wang–led restructuring. Meta also said it is opening ad accounts to AI agents, a development that will reshape how brands and PR teams manage paid social.
AI News Roundup
Substack Highlights
Substack-specific tooling is not currently connected for direct inbox reads, so today’s highlights are pulled from the Substack feeds Cam subscribes to via Inoreader. The full set of 143 newsletter subscriptions will be covered automatically once the Substack connector is restored.
Inoreader AI Folder
Gemini & New Evidence That NARA’s Copperplate Printed Both the Vellum & Paper Stone Declaration of Independence Copies
Kalev Leetaru uses Google Gemini to compare high-resolution scans of the U.S. Declaration of Independence’s vellum and paper stone copies and finds visual evidence the same NARA copperplate produced both. A small but vivid demonstration of using a frontier model for genuine archival/historical research.
Open Up: Anthropic launches Claude Security for Enterprise; Meta opens ad accounts to AI agents
Roundup of two big enterprise stories: Anthropic’s new security-focused tier of Claude for regulated industries, and Meta’s decision to allow third-party AI agents to log into and operate ad accounts. Plus Perplexity launching Financial Research Tools.
Exponential View #572: When certainty is the real AI risk
Azeem Azhar’s dispatch from his China visit, with on-the-ground notes from Zhipu, MiniMax, Kimi, Alibaba, Xiaomi and ByteDance. The thesis: leaders’ over-confidence in AI outputs is now a bigger operational risk than model failure itself.
Tokenmaxxing: “Quanto você gastou em tokens?”, pergunta o CEO dos tokens
Portuguese translation (by Rodrigo Ghedin for Manual do Usuário) of Gerard’s “Tokenmaxxing” piece — a sharp-tongued critique of how AI vendors are bragging about token-spend rather than business value.
OpenClaw runs on 4 tools. That’s it.
Behind-the-scenes look at OpenClaw’s tiny “Pi” coding agent — only four tools, but it produces strong results. The argument: most agent designs are massively over-engineered and a focused toolset beats a sprawling one.
Anthropic’s Jupiter model leaks as it races for $900B valuation in two weeks
Reports of a leaked next-gen Anthropic model code-named “Jupiter,” timed against the company’s reported sprint toward a $900 billion valuation. Pairs with Google’s $40B follow-on commitment.
Lore Issue #182: Elon’s Legal Battle With OpenAI Begins
Opening arguments in Musk vs. OpenAI; the OpenAI–Microsoft alliance reset; Gemini turning chat into full office software; and China blocking Meta’s Manus deal. A useful single-stop digest of the weekend’s biggest corporate AI moves.
Week in AI: Big Tech is spending $725B and supply is still the bottleneck
Microsoft, Alphabet, Amazon and Meta all told Q1 investors that capacity (not demand) is what is holding back AI growth. The newsletter pulls out the procurement implications: enterprise compute deals are now being negotiated years in advance.
400,000 new devs? Why the “coding is dead” era was a myth
Pushes back on the “AI is killing coding” narrative with hiring and signup data showing the opposite — Codex, Claude Code and similar tools are bringing more new entrants into software development, not fewer. Bonus: an unusual ChatGPT obsession story.
No-win? (Plus, OpenAI gets sued for > $1B)
Matt’s roundup leads with a billion-dollar-plus lawsuit against OpenAI alongside short reviews of new tools. Useful for keeping a running tally of AI legal exposure as it accumulates.
Humanoid robots are entering real production
Tracks several humanoid robotics programs moving from prototype to small-batch production this quarter. Includes a separate item on an AI tool that detects pancreatic cancer years earlier than current clinical workflows.
“AI bubble: a multi-company Enron waiting to happen” — David Gerard on Stupid Sexy Privacy
47-minute podcast interview by BJ Mendelson with Gerard, who lays out his case that the current AI investment cycle has the structural fingerprints of a multi-company accounting blowup. Useful counter-narrative material for any media-prep work.
Kim Crawley: Technofascism Survival Guide
Podcast interview with Kim Crawley of StopGenAI, covering her forthcoming book on resisting AI-driven authoritarianism. Heavily editorialised, but useful to know what the most active critic camp is publishing.
OpenAI to Microsoft: It’s not me, it’s you
Plain-English breakdown of how the OpenAI–Microsoft 2019 deal worked, why both sides chose to revise it this week, and what the new arrangement means for OpenAI’s distribution choices going forward.
OpenAI introduces Advanced Account Security
OpenAI rolls out phishing-resistant login, stronger account recovery, and protections against account takeover. Notable for any organisation running ChatGPT Edu or Business accounts at scale — including any team standardising on ChatGPT for sensitive comms work.
LIVE NOW: Learn To Make Agents In ChatGPT
Live training on building custom agents inside ChatGPT, aimed at Edu/Biz account holders. Worth flagging if any internal teams are exploring ChatGPT-based agents for content or media-monitoring workflows.
AI Workflows & Tool Watch
MacWhisper 13.20 ships full command-line tool (“mw”)
MacWhisper now includes an official CLI that lets you transcribe and stream files using any local or cloud Whisper model, switch models from the terminal, pipe transcripts into other programs, and call LLMs to clean up output before passing it on to the next tool. In plain English: you can now wire MacWhisper directly into Hazel, Keyboard Maestro, Drafts actions, or n8n workflows — no Apple Script gymnastics required. For podcast post-production this changes the math on automated editing.
DEVONthink 4.2.2 “Cassini” + DEVONthink To Go 4.0.6 — refreshed AI integration
The latest DEVONthink updates support ChatGPT, Claude, Gemini, Mistral and Perplexity natively, and Joe Kissell’s Take Control of DEVONthink 4 has been rewritten to match. For knowledge-management workflows this means you can now have any of those models summarise, classify, and search inside a DEVONthink database without bouncing through a browser. Pairs naturally with the MacWhisper CLI for an end-to-end “transcribe → file → summarise” pipeline.
Claude Cowork now runs your Obsidian vault as a knowledge engine
Now that Claude Cowork is generally available, third-party guides are publishing detailed playbooks for using it as the automation layer over an Obsidian vault — auto-generating daily notes, web-clipping, doing backlink completion across thousands of files, and even writing plugins on demand. Drop a CLAUDE.md file at the root of the vault and Cowork picks up your project conventions automatically. This is the closest thing to a “second brain that runs itself” Cam has flagged interest in.
Self-building n8n workflows via Claude Code
A pattern showing up in multiple write-ups this week: instead of clicking together n8n automations by hand, you describe the workflow in plain English to Claude Code, which writes, tests and deploys the n8n flow for you. Ability.ai and MindStudio both report build time dropping from “hours” to “minutes” on real client work. One detailed case study covers a small business using a multi-agent n8n stack to triage WhatsApp messages and voice calls into four sub-agents: booking, quoting, stock-check, and human escalation. Same pattern would apply to inbound media inquiries.
New MCP servers to know about
Three notable Model Context Protocol releases this week: Omnea (procurement data — first intake/orchestration platform with native MCP across Claude, ChatGPT, Microsoft Copilot, Cohere North and Cursor); Optimizely (experimentation — exposes flag management and experiment configuration as chat-callable actions); and TheBrain (a network-style notes app — full integration with Claude, ChatGPT and others). Plus a community “ha-mcp” connector for Home Assistant has hit the front page of the HA forum. MCP, in plain language, is the universal adapter that lets Claude or ChatGPT “see” and act on data inside another app.
“MCP Apps” — interactive HTML inside chat
A subtle but important shift co-developed by Anthropic and OpenAI: until January, every MCP interaction was text-only. MCP Apps now lets a tool return rich, interactive HTML that renders inside a sandboxed iframe right in your chat — dashboards you can manipulate, designs you can edit, formatted messages you can compose. Works in Claude, ChatGPT, Goose and VS Code. For comms teams this is the moment the chat window becomes a real workspace rather than a Q&A box.
OpenClaw’s “less is more” agent design (4 tools)
The Neuron’s deep-dive on OpenClaw’s coding agent “Pi” — which runs on only four tools and consistently beats more elaborate agent rigs. The takeaway for anyone designing internal AI workflows: a small, well-chosen tool palette plus clear instructions outperforms a sprawling “everything” agent. A useful reference if anyone at Tencent is scoping in-house agent toolkits this quarter.
Perplexity Comet rolls out to enterprises
Perplexity’s AI-native browser, Comet, now ships with silent enterprise deployment via MDM (Mobile Device Management) for both macOS and Windows. The “Comet Assistant” can do in-page summarisation and run multi-step actions like booking flights, managing email and filling forms. For a global comms team that travels heavily, the booking/email handling alone is worth a closer look.
Anthropic’s Blender connector for Claude
The Blender developers shipped an official MCP connector that lets Claude analyse and debug entire 3D scenes, or run scripts that batch-apply changes to objects in a scene. Niche, but a useful proof-point that the MCP connector ecosystem is now reaching specialised creative tooling — not just productivity SaaS.
Tencent Mentions
HSBC: Tencent’s Hunyuan 3 preview narrows the gap with top Chinese models on coding & agents
HSBC Research note over the weekend (via AAStocks) calls out HY3 as a real step-change in Tencent’s agentic and coding capability. Worth tracking how Western press picks this up — it is the strongest pro-Tencent AI line of the past 30 days.
Tencent Cloud upgrades full-stack enterprise Agent capabilities; expands Liangjiang (Chongqing) partnership
At the 2026 Tencent Cloud City Summit, Tencent Cloud rolled out ClawPro Private Cloud Edition, ADP Intelligent Workbench, Agent Memory and Agent Storage, and announced a deeper partnership with Liangjiang New District in Chongqing (AAStocks).
Tencent Cloud’s third 2026 price hike: AI coding assistant up 154%
The CodeBuddy enterprise flagship rises from 78 yuan to 198 yuan per user/month effective May 15; the dedicated edition doubles from 158 to 316 yuan. Expect this to be picked up by international tech press as evidence of broader compute-scarcity inflation in China.
CMSI: DeepSeek V4 refresh names Tencent and Alibaba as direct beneficiaries
China Merchants Securities International argues both Tencent and Alibaba benefit directly because they host and commercialise DeepSeek models on TokenHub and Bailian respectively. AAStocks coverage notes the agent-track competition is intensifying.
Tencent Cloud launches DeepSeek-V4 on TokenHub
TokenHub now serves DeepSeek-V4 directly to enterprise customers — a tangible follow-through on the “Tencent benefits from open-source SOTA” thesis above.
Yahoo Finance: Tencent Music (TME) narrative shifts on mixed analyst signals
Standalone TME story; Simply Wall St also asks whether the recent share-price weakness has created a value entry. Not AI-led, but worth flagging given how often Tencent Music gets used as a proxy for Tencent’s broader media business.