AI Intelligence Report
Top Stories
Anthropic doubles Claude rate limits after striking a compute deal with SpaceX’s Colossus 1
Anthropic announced it has signed a deal to use the entire compute capacity of SpaceX/xAI’s Colossus 1 data center in Memphis — more than 300 megawatts and over 220,000 NVIDIA GPUs (H100, H200, and next-generation GB200 chips). On the back of that capacity, Claude Code’s five-hour rate limits have been doubled, peak-hour throttling has been removed for Pro and Max plans, and Claude Opus API rate limits have jumped by more than tenfold (tier 1 input tokens went from 30,000 to 500,000 per minute). Anthropic also said it is exploring “orbital” AI compute with SpaceX. The deal arrives just weeks before SpaceX’s planned IPO and reshapes the AI infrastructure map: Anthropic now has access to compute originally built for Elon Musk’s xAI.
EU Council and Parliament reach political deal to simplify the AI Act ahead of August deadline
EU lawmakers reached a provisional agreement on May 7 to streamline parts of the AI Act before its main provisions take effect on August 2, 2026. National AI regulatory sandboxes have been delayed by a year (now August 2027), and — most relevant for communications and content teams — the grace period for providers to implement transparency labelling for AI-generated content has been cut from six months to three months, with a hard deadline of December 2, 2026. The simplification is being framed as helping European AI providers compete, but it tightens the timetable for AI-generated media disclosure.
OpenAI launches GPT-5.5-Cyber, a security-focused model for vulnerability research
OpenAI expanded its Trusted Access for Cyber programme with two new models — GPT-5.5 and GPT-5.5-Cyber — aimed at verified defenders working on vulnerability research and critical-infrastructure protection. The release follows a wider OpenAI push this week that also included a new Trusted Contact safety feature in ChatGPT (which can alert a chosen friend or family member if the system detects serious self-harm risk) and a refreshed set of realtime voice models in the API. Expect AI-assisted security research to become a much louder talking point in the coming weeks.
Cloudflare announces 20% workforce cut as it pivots to an “agentic AI-first” operating model
Cloudflare reported Q1 revenue of $639.8 million (up 34% year over year) but said it will reduce headcount by roughly 20% and book up to $150 million in restructuring charges as it shifts to what CEO Matthew Prince described as an agentic-AI-first operating model. It is one of the largest AI-driven layoffs at a major US public tech company so far this year and a leading indicator of how mainstream tech firms are restructuring around autonomous agents — a useful talking point for any internal-comms conversations about AI’s effect on workforce planning.
Backlash builds as Chrome quietly force-installs a 4GB Gemini Nano AI model
Users this week noticed that Google Chrome silently downloads a roughly 4GB on-device AI model (Gemini Nano) on every desktop install. The model has technically been there since 2024, but it became a talking point this week as users discovered it eating disk space without notice. Coverage is critical of the lack of opt-in. For comms teams this is a useful reference point on what “AI by default” backlash looks like and how it can be avoided.
Trump administration’s CAISI strikes pre-release AI model evaluation deals with Google DeepMind, Microsoft and xAI
The Center for AI Standards and Innovation — the renamed body inside NIST — has struck agreements with Google DeepMind, Microsoft and xAI that allow the U.S. government to evaluate frontier AI models before they are released to the public. It also renegotiates earlier pre-release access deals with OpenAI and Anthropic. This is now the closest thing the U.S. has to mandatory pre-deployment safety review for the largest model providers, and it is a significant development for any communications team that deals with regulators or government affairs.
AI News Roundup
Substack & Newsletter Highlights
Drawn from the AI-focused Substacks and email newsletters that arrived in the past 24 hours. Direct Substack inbox access was unavailable for this run, so summaries are sourced via the corresponding RSS feeds in your Inoreader AI folder.
Inoreader AI Folder
Articles from the past ~24 hours in your Inoreader “AI” folder, deduplicated against Top Stories and Substack Highlights.
Google Chrome force-installs a 4-gigabyte AI model — how to get rid of it
Walks through how Chrome’s silent 4GB Gemini Nano download works, what it’s for, and how desktop users can disable or remove it. Useful as a primer if anyone on your team asks “is Chrome doing AI on me without telling me?”
Scaling Trusted Access for Cyber with GPT-5.5 and GPT-5.5-Cyber
OpenAI’s official write-up of how GPT-5.5-Cyber works and how verified defenders can apply for access. Worth a skim before any conversation about responsible AI release in security contexts.
Introducing Trusted Contact in ChatGPT
OpenAI rolled out an opt-in Trusted Contact feature: a designated person can be alerted if the system detects serious risk of self-harm in the conversation. A potentially significant move for AI safety messaging — especially given how regulators are increasingly focused on chatbot duty-of-care.
Advancing voice intelligence with new models in the API
OpenAI announced a new realtime voice model line for the API that can reason, translate, and transcribe speech in close to real-time. Relevant to anyone planning AI-driven podcast or interview tooling.
Parloa builds service agents customers want to talk to
Case study on Parloa, which uses OpenAI to power voice-driven customer-service agents and lets enterprises simulate agent behaviour before deployment. A useful reference example for AI customer-service positioning.
Simplex rethinks software development with Codex
Simplex says it has compressed design, build and test cycles using ChatGPT Enterprise and Codex; another data point for OpenAI’s “AI as workforce” pitch.
Uber uses OpenAI to help people earn smarter and book faster
OpenAI-powered AI assistants and voice features now sit inside Uber’s app for both drivers and riders, with a focus on real-time, marketplace-aware decisions.
Singular Bank helps bankers move fast with ChatGPT and Codex
Spanish private bank built “Singularity,” an internal AI assistant that reportedly saves bankers 60–90 minutes a day on meeting prep, portfolio analysis, and follow-ups.
How frontier enterprises are building an AI advantage
OpenAI’s “B2B Signals” research argues that the difference between leading and trailing enterprises is now the depth of how AI is woven into workflows — not whether AI is available. Top performers use AI roughly 3.5× more intensively.
Introducing ChatGPT Futures: Class of 2026
OpenAI named 26 student “Futures” innovators using ChatGPT to build, research, and ship projects. Reads as a brand and recruiting play aimed at universities.
OpenAI ChatGPT goes goblin mode — let none say ‘model collapse’
Critical take on persistent oddities in GPT-5.5 outputs (the “goblin mode” anecdotes) and what they suggest about training-data quality. Worth bookmarking as ammunition for the “is AI getting better or worse?” debate.
90% of AI tools are noise. Here are the 10%.
Live episode promo arguing that the vast majority of AI products in 2026 are derivative wrappers, with a curated shortlist of those actually worth installing. Useful for the workflows-and-tools section of your team’s reading list.
Watch: $6B per drug. 90% fail. This team has a fix.
Profile of an AI-driven drug-discovery team and how Google’s protein-folding work is being applied. Niche but a good “AI is changing science” anecdote for talking points.
AI Workflows & Tool Watch
Claude Code rate limits doubled, peak-hour throttling removed
Anthropic’s Pro and Max plans now have double the five-hour Claude Code limit and no longer slow down at peak times. If you canned Claude Code in April because of throttling, today is the day to give it another spin. Claude Opus API limits jumped roughly 16× for tier-1 users — meaningful headroom for batch tasks.
n8n-mcp: build n8n workflows directly from Claude Code or Cursor
The n8n-mcp project has matured into a real production tool. It exposes n8n as a Model Context Protocol server, so Claude Code (or Cursor, Claude Desktop, Windsurf) can read and write directly into your n8n instance. Describe an automation in plain English and Claude will architect the nodes, triggers, and data transformations — a 20-minute setup according to early adopters. This is exactly the “Claude as builder” story that fits how you already use n8n for podcasting and PR pipelines.
Obsidian “second brain” + Claude Code skills
A new generation of Claude skills wraps Obsidian-specific conventions: automatic tagging, cross-linking, vault structure awareness, and template-aware note creation. The result is that Claude stops behaving like a generic assistant and starts treating your Obsidian vault as your personal knowledge graph. Worth pairing with your existing Drafts and Apple Notes workflow.
Perplexity Computer arrives in Microsoft Teams
Perplexity Computer — its agentic research/browse/document tool — is now an installable Microsoft Teams app. Teams users can orchestrate research, analysis, and document creation directly inside Teams conversations. For organizations standardized on Teams, this is the most direct way yet to put a Perplexity-style agent in front of comms staff without forcing a separate app.
Microsoft Agent 365 + Copilot Cowork
Microsoft is positioning “Agent 365” as the IT-admin console for managing AI agents at scale across Microsoft 365, while Copilot Cowork moves Copilot from a sidebar helper to a delegate that can carry out multi-step tasks. If your team is on Microsoft 365, this is the operating layer your IT department is about to be evaluating.
Voice agents for newsroom and PR work
Two announcements this week — OpenAI’s new realtime voice models and Parloa’s voice-agent platform — bring legitimate quality to voice-AI applications. Practical first-use cases for a comms team: real-time interview transcription beyond MacWhisper, multilingual press-conference live translation, and “rehearsal” simulations of difficult media calls. The new OpenAI voice models can reason, translate, and transcribe in one pass.
Product Hunt launches worth a look
This week’s notable Product Hunt AI launches: Gas City 1.0 (multi-agent CI/CD orchestration), Tollecode (local-first AI coding agent emphasising safe shell execution and model switching), Git Pitcher (turns a repo into a structured plan and prompt pack for downstream agents), FlowMarket (B2B deal matching/scheduling agents), and Basedash (governed company data into conversational analytics). The clear theme: agents moving from general assistants to specialised operators.
Claude Code’s MCP improvements (changelog highlights)
Recent Claude Code updates worth noting: the /mcp command now shows tool counts and flags servers that connected with zero tools; failed MCP server starts now auto-retry up to three times; reconnects no longer flood the conversation with full tool-name lists; and MCP Tool Search lazily loads tool schemas, cutting context use by up to 95% on big setups. If you’ve been hitting context-window pain on Claude Code, the Tool Search behaviour is a meaningful upgrade.
Tencent Mentions
Tencent rebound masks AI doubts ahead of May 13 Q1 results
Coverage in Hong Kong financial press argues that the recent share price recovery is masking analyst concern about whether Tencent’s AI execution can keep pace with Alibaba and ByteDance. Q1 results land Tuesday, May 13 — analysts will be looking specifically for Hunyuan iteration cadence, WeChat-native AI agents, and clearer guidance on capital returns. Worth pre-briefing for media calls next week.
Hy3 preview model — performance positioning
External coverage continues to highlight Hy3 preview as a 295-billion-parameter MoE model with 21 billion active parameters and a 256K context window, framed as a fast-and-slow-thinking fused model for reasoning, coding, and agentic workloads. International analysts note this is the first major model to emerge from Tencent’s post-February pre-training and RL infrastructure rebuild.
Tencent doubles down on agentic AI with latest Hunyuan updates
KrAsia frames Tencent as positioning Hunyuan firmly around agentic capability rather than raw scale — an explicit contrast to U.S. labs. The article cites the reported plan to roughly double AI investment in 2026 to over RMB36 billion, up from ~RMB18 billion in 2025.
Global rollout of scenario-based AI capabilities
Tencent’s official channel highlighted the global rollout of “scenario-based” AI capabilities aimed at industrial efficiency. Useful framing if international press asks about Tencent’s overseas AI strategy.
Former OpenAI research scientist launches new AI model for Tencent
InfoWorld reporting on the rebuild of Tencent’s AI research bench, including the recruitment of a former OpenAI research scientist as a key technical lead behind Hy3. Worth noting as a “talent flow back to China” story that international tech media will likely revisit.