AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
This was the week AI became as much about control as capability. The US government reportedly forced Anthropic to take Claude Mythos Preview and Claude Fable 5 offline for foreign users, citing national-security concerns. Anthropic disabled both models globally, while Commerce Secretary Howard Lutnick said the company would need a licence to make them available to “foreign persons”, including its own overseas employees. It is a reminder that frontier models are no longer just products. They are now geopolitical assets.
Markets were just as dramatic. SpaceX listed and closed at a $1.77tn valuation, while Cursor owner Anysphere raised $60m at a $9.6bn valuation after crossing $2bn in annualised revenue. OpenAI, by contrast, burned $3.7bn in Q1 on $5.7bn of revenue, ending the quarter with more than $73bn in cash. The capital gap between the biggest labs and everyone else is widening, but it also gives OpenAI and Anthropic more room to delay IPOs if they choose to.
The practical message for UK leaders is clearer than the headlines. First, if your team has a real workflow around Claude Mythos Preview or Claude Fable 5, swap the endpoint for OpenRouter Fusion and test performance before Friday. Second, do not start AI procurement with the model. Start with the work: one document, one contract, one repeatable process. Third, if you spend more than £500 a month on agentic coding, run one heavy-use repo through Cursor Composer 2.5 fast mode for a week and compare the result against your current setup.
What to Try This Week: Schedule a Morning Briefing in ChatGPT
OpenAI added a Scheduled Tasks tab in ChatGPT so you can run recurring prompts without retyping them. The easy win is a daily briefing that saves you about ten minutes every morning.
Try this:
Open ChatGPT on web or mobile
Go to the new Scheduled tab
Click Create Task
Add a prompt such as “Every weekday at 6am, send me a concise AI news briefing covering model releases, funding news, product launches and anything relevant to strategy, software, M&A or productivity.”
Choose Daily, set the time, and save
Manage or edit it anytime from the Scheduled page
Geopolitics, Governance and Big Moves
Anthropic’s Fable 5 Pause Turns Frontier AI Into a Policy Test Case
Anthropic has taken Claude Fable 5 offline after a US government order linked to alleged jailbreak and national security concerns. The model, launched days earlier for long-running autonomous coding and knowledge work, had been positioned as one of Anthropic’s most capable agents. Anthropic says it complied while disputing the severity of the issue, turning Fable 5 into a live test of how frontier models may be regulated under pressure globally.
ChatGPT Scheduled Tasks Move AI From Assistant to Operating Rhythm
OpenAI has made Scheduled Tasks easier to find and manage, giving users a dedicated place to create reminders, recurring prompts and daily briefings. Instead of treating ChatGPT as a one-off answer machine, users can now ask it to run prompts later, daily, weekly or on a custom cadence. It supports practical loops like AI news updates, market monitoring, meeting prep and research summaries, with active-task limits varying by plan tier.
SpaceX Listing Becomes an AI Infrastructure Raise in Disguise
SPCX raised $75B at $135 a share and jumped on debut, while Senator Warren asked the SEC to delay the offer. SpaceX also confirmed it will exercise its option to acquire Cursor. The framing: the listing is an AI infrastructure fundraise, with SpaceX positioned to launch AI compute and software into space.sly.
OpenAI - Burned $3.7B in Q1, Still Has $73B Cash
OpenAI’s Q1 burn was around $3.7B on $5.7B revenue, but it ended the quarter with roughly $73B in cash and securities after a major March funding round. At that burn rate the IPO pressure is lower, and the filing becomes more about optionality than urgency.
Dario Amodei Pushes a Frontier-Lab Regulator With Model Grounding Powers
Amodei published “Policy on the AI Exponential”, calling for an FAA-style regulator able to ground frontier models that fail safety tests, plus labour buffers like wage insurance. The document explicitly flags hacking capability as strategically consequential, and landed just before the export-control directive.
Same-Day AI Liability Rulings Hit Google’s AI Overviews and Platform Design
A German court ruled Google is responsible for false statements in AI Overviews, rejecting the idea users can just check links. In California, Section 230 was held not to shield design choices in a young-user-harm case. The pattern: new constraints on AI outputs and product design.
Jeff Bezos pitches Prometheus as the “artificial general engineer” at a $41 billion valuation
Bezos disclosed a $12B Series B at a $41B valuation for Prometheus, pitching an “artificial general engineer” to compress the dream-to-build loop by 10x for complex machines. He also argued AI productivity creates more opportunity, not less, positioning the company against the dominant “AI freezes hiring” narrative.
DeepSeek’s $7.4bn raise turns efficient AI into China’s national champion story
DeepSeek has reportedly closed its first major external funding round, raising more than $7.4bn at a valuation above $50bn. The round is striking not just for its size, but for its structure: most investors are backing a limited partnership controlled by founder Liang Wenfeng, rather than taking direct voting power in DeepSeek. That lets the company fund compute, research and commercial products while preserving founder control. The bigger signal is geopolitical. China’s most visible open-weight AI lab now has serious financial firepower behind its efficiency-first model strategy, even as US chip restrictions bite.
Tech, Tools and Releases
GLM-5.2 Raises the Bar for Open-Weights Coding Models
Z.ai’s GLM-5.2 is the week’s standout open-weights release: a MIT-licensed, text-only MoE model with around 744–753B parameters, 40B active, and a claimed 1M-token context window. Built for repository-scale coding agents, it leads open models on long-horizon software benchmarks and, on Z.ai’s published figures, edges GPT-5.5 on FrontierSWE while sitting just behind Claude Opus 4.8. Accessible via Hugging Face and compatible APIs, it narrows the open/proprietary gap in practice this week.
Moonshot’s Kimi K2.7-Code raises the bar for open-weight coding agents
Moonshot AI has released Kimi K2.7-Code, a coding-focused open-weight model built on Kimi K2.6. The model keeps the 1 trillion-parameter MoE architecture but is tuned for real-world, long-horizon software engineering tasks: planning, editing, debugging and agentic workflows across large codebases. The headline improvement is efficiency. Moonshot says K2.7-Code uses around 30% fewer thinking tokens than K2.6, which matters because agentic coding costs can balloon quickly. It also supports long context, vision and tool use, making it another serious Chinese open-weight challenger in the developer-tools race.
Perplexity Gives Its Computer Agent a Memory That Learns From Work
Perplexity has launched Brain, a Research Preview memory layer for its Computer agent, available to Max and Enterprise Max users. Rather than saving preferences, Brain builds a traceable work memory across projects, files, connectors and past sessions, so the agent can reuse what worked and avoid repeating mistakes. Perplexity says early tests show 25% better correctness, 16% higher recall and 13% lower cost on repeated, context-heavy, recurring agent work loops.
OpenRouter - Launches Fusion for Model-Panel “Neurodiverse” Answers
OpenRouter’s Fusion pools multiple models, has a separate model evaluate each, then merges into one reply. A reported panel narrowly trailed a benchmark score attributed to Fable 5 at roughly half the cost. The pitch: the future is multiple models working together rather than one takeover model.
Cursor Composer 2.5 lands as a specialist counter to generalist coding models
Composer 2.5 is Cursor’s native model, RL-fine-tuned in a simulated CLI harness. The claim is strong cost-speed performance versus generalist options on certain tasks, supporting the thesis that specialist models can win on dollars and seconds even if generalists win on broad benchmarks.
Open-source agent harnesses arrive at scale: OpenCoworker, MiMo Code, SkillSpector
OpenCoworker landed as a desktop-agent harness alternative to closed desktop agents. MiMo Code shipped as a terminal-native coding assistant with a memory system, and SkillSpector scans agent skills for vulnerabilities before installation. The shared driver is procurement risk if policies and retention change overnight.
Anthropic - Ships Claude for Word Inside Microsoft Word
Anthropic now ships a first-party Microsoft Word integration that puts Claude in a sidebar, so users can rewrite or restructure a document without leaving Word. Available on all paid Claude plans. The setup is five clicks: install the add-in, click the Claude icon in the upper-right corner, log in, ask for a rewrite / summary / targeted edit, follow up iteratively. Why now matters: Microsoft is internally restricting Claude Fable 5 from its own GitHub Copilot pending a legal review of Anthropic’s new 30-day data-retention regime — which makes the consumer Claude-for-Word integration the cleanest way for non-Microsoft teams to get Fable-class quality on Word documents without waiting for GitHub Copilot to enable it.
Visa plugs into ChatGPT for agentic checkout
Visa connected its network to ChatGPT’s agent system so agents can execute transactions with user-defined caps, merchant restrictions, and approvals, with Visa managing fraud and chargebacks. It is framed as a new payment rail, with initial use cases including shopping, invoices, and coding agents buying APIs.
Databricks launches Omnigent: the control layer for multi-agent work
Databricks has open-sourced Omnigent, an Apache 2.0 “meta-harness” designed to sit above tools such as Claude Code, Codex, Pi and custom agents. The idea is simple: as teams start using multiple AI agents, the pain shifts from prompting to orchestration. Omnigent provides a shared layer for composing agents, switching models, enforcing policies, controlling spend, pausing risky actions, running agents in sandboxes and collaborating on live sessions. It is still alpha, but the direction is important. The agent market may not standardise around one tool, but around a control layer above them.
Quick Hits
Meta’s tokenmaxxing-to-tokenminimizing reversal. In April Meta CTO Andrew Bosworth told Forbes employees should “Keep doing it. No limit” on AI token usage. By June, Meta has told staffers in a memo it plans to impose limits on AI token usage to control costs running into the billions. Pair with the Rivos AI-chip failure scoop and Beijing’s $2 billion Manus deal unwinding for the full picture: external AI assets cut off, internal AI structure in revolt, in-house chip programme not yet working.
Anthropic’s Max-plan class action. A Washington-based subscriber who upgraded to Anthropic’s $200-per-month Max 20x plan filed Sunday night, alleging usage limits sold as 5× and 20× the Pro tier were in practice “significantly lower.” Lands the same day Anthropic dispatched engineers to the White House.
100+ cybersecurity executives sign the Free Fable open letter. Names tied to Adobe, Zoom, Sophos, Vercel, Veracode, Nvidia and Stanford HAI signed; ex-Facebook security head Alex Stamos called the flagged jailbreak “a proof of concept of a flaw” - the same kind defensive teams use to patch weak spots - and the letter named OpenAI’s Daybreak, GPT-5.5, Kimi 2.7, Opus and Sonnet as all having the same flaw-finding capability.
FIFA World Cup 2026 opens with AI wired into every layer. 150 million optical-tracking data points per match, an Adidas ball reporting 500 times a second, 3D body scans for every player, and a Football AI Pro chatbot for all 48 squads. Google made Gemini the global sponsor of defending-champion Argentina - first frontier-chatbot brand to sponsor a national football team.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.