AI Intelligence Report
Top Stories
Google to pour up to $40 billion into Anthropic, deepening AI alliance
Alphabet has committed an immediate $10 billion at a $350 billion valuation, with another $30 billion contingent on Anthropic hitting performance milestones. The deal also bundles a multi-gigawatt expansion of TPU chip capacity through Broadcom, beginning in 2027. The investment signals that the era of a clear AI “Big Three” is fracturing — Google is hedging by simultaneously building its own Gemini line and bankrolling its closest rival, while OpenAI is increasingly isolated. The Motley Fool argues Google has likely secured a bargain at this valuation given Anthropic’s reported $30 billion ARR (a 30-fold jump in a single year).
Anthropic launches Project Glasswing with secret “Mythos” model — finds thousands of zero-day flaws
Anthropic unveiled Project Glasswing, a coalition with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, the Linux Foundation and 40+ other organizations. The initiative gives partners controlled access to Claude Mythos Preview, an unreleased frontier model, for defensive cybersecurity work. In just weeks Mythos has autonomously identified thousands of zero-days across major operating systems and browsers — including a 17-year-old root-access flaw in FreeBSD’s NFS implementation, a 16-year-old FFmpeg vulnerability, and a 27-year-old OpenBSD bug. Anthropic is committing $100 million in model credits to the program but explicitly will not release Mythos to the general public. Independent researcher Simon Willison called the restricted release “necessary.”
DeepSeek V4 lands with 1M-token context, rock-bottom prices, and Huawei chip integration
Hangzhou-based DeepSeek released a preview of V4 on Friday in two flavours: V4 Flash ($0.14 / $0.28 per million input/output tokens) and V4 Pro ($0.145 / $3.48). Both feature a 1 million token context window — an eight-fold jump over V3 — and major upgrades to reasoning and agentic abilities. Crucially, the model was trained and runs on Huawei “Ascend 950” chips clustered via Huawei’s “Supernode” technology, marking a meaningful step toward a Chinese AI stack independent of NVIDIA. Like prior DeepSeek releases, V4 is open source. MIT Technology Review noted the launch is unlikely to repeat last year’s market shock since investors have already absorbed the reality that Chinese AI is competitive.
Tencent unveils Hunyuan Hy3 Preview — first model since the Hunyuan rebuild
Tencent released the Hy3 Preview, a 295-billion-parameter “fast-and-slow-thinking fused” mixture-of-experts model with 21 billion activated parameters and 256K context. The model was already deployed inside Yuanbao, CodeBuddy, WorkBuddy, ima, Tencent Docs and Peacekeeper Elite before public launch. Benchmarks: SWE-bench Verified 74.4%, Terminal-Bench 2.0 54.4%, BrowseComp 67.1%, WideSearch 70.2% — competitive with GLM-5 and closing meaningful ground on Claude Opus 4.6 and GPT-5.4. Pricing on Tencent Cloud TokenHub: RMB 1.2 per million input tokens, RMB 4 per million output tokens, with a two-week free token window. The launch came less than three months after Tencent reset its pre-training and reinforcement-learning infrastructure — a notably fast turnaround.
OpenAI and Microsoft restructure their partnership — OpenAI freed to use any cloud
The two companies announced an amended agreement that lets OpenAI serve all its products across any cloud provider. Microsoft will no longer collect a revenue share from OpenAI. The deal adds long-term clarity to a relationship that had grown strained as OpenAI courted Oracle, Google Cloud and others for compute capacity. Same week, Sam Altman published “Our Principles,” a five-point statement of OpenAI’s values that several commentators (including The Neuron) noted arrived “one day too late” given the simultaneous Microsoft restructuring news.
AI News Roundup
Inoreader AI Folder & Newsletter Inbox
OpenAI available at FedRAMP Moderate
OpenAI announced that ChatGPT Enterprise and the OpenAI API now hold FedRAMP Moderate authorization, the U.S. government’s standard for storing controlled-but-unclassified federal data. This unlocks federal-agency adoption that was previously blocked on procurement grounds.
The next phase of the Microsoft–OpenAI partnership
OpenAI and Microsoft formally announced their amended agreement: OpenAI can now serve all products on any cloud, and Microsoft no longer takes a revenue share. The release frames it as adding “long-term clarity” — language that telegraphs the strain that has built up over compute-supply disagreements.
An open-source spec for orchestration: Symphony
OpenAI released Symphony, an open spec for orchestrating Codex agents around issue trackers (think Jira, Linear, GitHub Issues). Idea: rather than humans dispatching agents, the issue tracker itself becomes the always-on dispatcher, reducing context-switching for engineering teams.
Choco automates food distribution with AI agents
OpenAI customer story: Berlin-based Choco used the OpenAI API to streamline B2B food distribution, citing productivity gains and growth unlocked by AI agents managing routine ordering workflows.
Our Principles
Sam Altman lays out OpenAI’s five guiding principles for ensuring AGI benefits all of humanity. The commentary in The Neuron’s “Bad timing, Sam” newsletter points out the awkward overlap with the same week’s Microsoft restructuring announcement.
Signal in the Noise: Today’s Intelligence Advantage for Comms Leaders
Directly relevant to your role: Signal AI summarizes findings from recent industry reports on how communications leaders are turning AI-driven media intelligence into faster decisions, paired with practitioner perspectives from a recent webinar. Useful framing for the Tencent global comms team’s own monitoring stack.
Meta’s Manus acquisition blocked by China (PLUS: Google to invest $40B in Anthropic)
AI Valley’s daily roundup leads with Beijing blocking Meta’s purchase of Chinese AI-agent startup Manus — a notable cross-border M&A signal — and frames the Google–Anthropic megadeal as further evidence that the AI capital race is concentrating around two labs.
AI pricing collapses further as DeepSeek cuts cache costs by 90 percent
AI Breakfast covers DeepSeek’s cache-cost cut and walks through what it means for cumulative inference economics, particularly for builders using cached prompts at scale (RAG pipelines, customer-support bots, agentic tool chains).
😼 Bad timing, Sam
The Neuron’s daily skewers the awkward optics of OpenAI publishing “Our Principles” the same week as the Microsoft cloud-share restructuring. Sidebar items: AI compute is now costing more than salaries at some firms; DeepSeek’s price war is biting U.S. providers’ margins.
😺 One analyst replaced 100 economists
The Neuron profiles “Jeremy,” an analyst reportedly spending $6,000/day on Claude API calls and replacing the work of a 100-person research team. Frames Meta’s 10% workforce cut as the same dynamic at scale: capital being redirected from headcount to compute. The takeaway in the headline — “Don’t get replaced. Get leveraged” — is a useful lens for any function (including comms) thinking about AI-augmented productivity.
The Prompt That Builds Your First Online Course
A practical prompt template that turns a rough subject-matter idea into a structured course outline in minutes. Translatable to internal training materials, onboarding decks, or media-training curricula for spokespeople.
AI Agents are now shopping, haggling, and closing real deals!
Roundup of agent-driven commerce experiments now closing real transactions, plus a sidebar on prompt design for landing pages that actually convert. Notable for PR/marketing: agents are starting to be the buyer, not just the assistant.
Using Gemini 3 & Gemini Deep Research to triage the White House Correspondents’ Dinner shooting after 24 hours
GDELT’s Kalev Leetaru walks through how he used Gemini Deep Research to produce structured situational-awareness briefings on a fast-moving security incident within 24 hours. Highly relevant template for crisis-comms triage workflows — the same pattern works for any breaking issue where you need a synthesized, sourced backgrounder fast.
Using Gemini Deep Research to create “backgrounder” reports
Companion piece showing the same Gemini Deep Research technique applied to two complex topic areas (counterterrorism in India, crypto-for-terrorism-financing) to produce briefing-grade context documents — useful for internal pre-reads ahead of difficult interviews.
Deep trend analysis with Gemini on the Estonian International Security Service annual reviews 2021–2026
A worked example of multi-document, multi-year trend analysis using Gemini’s reasoning + visualization. Pattern is reusable: feed a model 5+ years of an organization’s annual reports and surface the shifts in language, priorities, and tone — the same approach works for tracking a competitor’s narrative arc over time.
🧙🏼 I ranked 21 industries by build-vs-buy
Dario at WhatPlugin.ai ranks 21 industries on whether they should build proprietary AI tools or rely on off-the-shelf platforms (and notes GPT-5.5 plus OpenAI’s new Workspace Agents in the same issue). A useful mental model for any internal AI investment decision.
Kohl-Verlag’s new line of AI-slop school textbooks
David Gerard surfaces Der Spiegel reporting on German publisher Kohl-Verlag, which is alleged to have produced AI-generated textbooks for inclusion classrooms. A cautionary tale for any organization shipping AI content into trust-sensitive contexts.
Claude Code rate limits: Anthropic squeezes the customers
David Gerard’s piece (translated to Portuguese this week) on Anthropic tightening Claude Code rate limits and the resulting power-user pushback. Worth tracking if any team in your org is building serious workflows on Claude Code.
AI Workflows & Tool Watch
Perplexity launches “Personal Computer” — turns a Mac mini into an always-on AI agent
Perplexity is rolling out Personal Computer for Mac, which runs continuously on a dedicated Mac mini and works directly with your local files, native macOS apps, and existing workflows — voice orchestration, file editing, browser actions and more. Comet Agent itself is now powered by Claude Sonnet 4.6 (Pro) or Opus 4.6 (Max), and Comet has expanded to Enterprise (silent MDM rollout, granular policy controls) and is now pre-orderable on iOS. Samsung’s Galaxy S26 also integrates Perplexity at the platform level.
Claude Code’s April updates land: 500K MCP results, parallel server connections, vim mode
The biggest practical win: Claude Code now connects to MCP servers in parallel at startup instead of one-by-one — no more long warm-up if you’ve connected several. MCP tool result persistence was raised from a much lower limit to 500K characters, meaning large database schemas, RSS dumps, or file listings no longer get truncated. Other ships: vim visual/visual-line modes, custom themes via JSON, /cost and /stats merged into /usage, PowerShell auto-approval, and a fix for a macOS keychain race that caused unexpected /login prompts.
The “Gemini Deep Research backgrounder” pattern for crisis & issues comms
GDELT’s Kalev Leetaru posted three back-to-back case studies this week showing how to use Gemini Deep Research to produce briefing-grade context documents within hours — for a breaking security incident (the WHCD shooting), for thematic research areas (counterterrorism in India), and for multi-year trend analysis (Estonia’s annual security reviews). The pattern: anchor the prompt in named source documents, ask for a structured output (timeline, key actors, open questions), and force citation. Direct lift for any global-comms team that needs to spin up a defensible backgrounder under deadline pressure.
OpenAI Symphony — issue tracker becomes the agent dispatcher
An open-source spec for Codex orchestration. Instead of you handing tasks to an agent one at a time, Symphony lets your issue tracker (Jira / Linear / GitHub) be the always-on dispatcher: tickets that match certain rules get picked up, worked, and returned with PRs/results. Most relevant to engineering teams, but the same pattern is portable to any structured task queue — including content production pipelines.
n8n + Obsidian community workflows worth a weekend project
The n8n community has been quietly publishing useful Obsidian-bridging templates: pulling Airtable data into Obsidian notes via AI summarization, converting an Obsidian vault into a private podcast feed read aloud by AI, and a local agentic AI that sorts fleeting notes overnight. The “Post Webhook” Obsidian plugin is the bridge most templates rely on. None of this requires giving up your file ownership — everything still lives in your local vault.
Signal AI’s “Intelligence Advantage” framing — for comms leaders
Signal AI’s latest practitioner write-up looks at how comms leaders are turning AI-driven media intelligence into faster, more decisive action — positioning AI not as a content-generation toy but as an early-warning and prioritization system. Usable as a reference point if you’re benchmarking the Tencent comms team’s own monitoring stack.
Tencent Mentions
Bloomberg: Tencent unveils Hunyuan Hy3 Preview in “high-stakes test” for its OpenAI hire
Bloomberg frames the Hy3 launch as a critical proof-of-life for the rebuilt Hunyuan team and the senior leader recruited from OpenAI to run it. The fast turnaround (under three months from infrastructure reset to public preview) is being read as confirmation that Tencent’s AI org is back on a competitive cadence.
Techi: Tencent Hy3 Preview — first model from the Hunyuan AI rebuild
Detailed walkthrough of Hy3 Preview’s architecture (295B total / 21B activated MoE), 256K context, and benchmark posture vs. GLM-5, Claude Opus 4.6 and GPT-5.4. Notes that the model was already in production inside Yuanbao, CodeBuddy, WorkBuddy, ima, Tencent Docs and Peacekeeper Elite before going public.
DaoInsights: “Hunyuan 3.0 — Tencent drops AI model at breakneck speed”
The China-tech focused outlet leads on the speed of execution and treats Hy3 as a signal that Tencent is now prepared to compete on cadence with the U.S. labs, not just on price.
Caixin: Tencent folds AI Lab into Hunyuan team in major AI overhaul
Background context that’s continuing to drive coverage: Tencent consolidated its AI Lab into the Hunyuan org last month, simplifying reporting lines and concentrating talent ahead of the Hy3 release.
Tencent.com: Global rollout of scenario-based AI capabilities
Tencent’s own announcement positioning its scenario-based AI offerings (industry-vertical templates, embedded copilots) for international customers. Useful talking-point material if international press asks how Tencent’s AI strategy differs from a pure foundation-model play.
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.