AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
This week’s developments show AI moving from models that answer questions towards systems that can complete work. OpenAI’s ChatGPT Work, Microsoft’s adoption of GPT-5.6 across Copilot and Google’s expansion of AI actions all point towards agents becoming embedded in everyday business processes. Meta is also beginning to charge developers directly for model access, while Moonshot’s Kimi K3 and Z.ai’s reported revenue growth underline the increasing commercial strength of China’s AI sector.
But capability is only half the story. Apple’s China launch arrangements, its legal dispute with OpenAI, proposed pre-release testing for frontier models and New York’s data-centre pause show regulation, intellectual property, energy and national sovereignty becoming central to AI strategy. Meanwhile, Europe’s funding market is increasingly concentrating around defence and sovereign technology.
For business leaders, the question is shifting from which model is best to which workflows can now be completed end to end, with suitable permissions, oversight and economics.
What to Try This Week
OpenAI has replaced its separate desktop experiences with a new ChatGPT app combining Chat, Work and Codex. Work is also rolling out across web and mobile to eligible paid accounts, so it may not appear for everyone immediately.
Once it appears, do not begin with an open-ended request to “analyse my business”. Choose one assignment you already understand well and for which you can judge the finished result. Good tests include preparing a customer-meeting pack, analysing a monthly budget variance or turning research and notes into a board-ready document.
Provide two or three genuine source files and, where appropriate, connect one relevant service such as Google Drive, SharePoint, email or your calendar. Ask Work to show you its plan before proceeding, specify the output format and identify anything that must not be changed. OpenAI recommends stating the required deliverable, source material, constraints and review criteria explicitly.
Geopolitics, Governance and Big Moves
Nobody can prove AI is killing jobs, and the economists admit it
The “We Must Act Now” statement, organised by Stanford’s Digital Economy Lab, warns AI will grow “radically more powerful” within a decade while conceding the profession cannot measure its effect on employment. Apollo’s Torsten Slok noted that five competing “AI exposure” frameworks disagree most precisely for the jobs thought most at risk. UVA’s Anton Korinek put it plainly: steam, electricity and computers each gave societies decades to adapt; AI may give only years. For workforce planning, plan under genuine uncertainty, not a forecast.
Hassabis proposes a 30-day safety gate for frontier AI
DeepMind chief Demis Hassabis has proposed an industry-funded US standards body, modelled on FINRA, to test frontier AI systems before release. Labs would initially submit models voluntarily for reviews lasting up to 30 days, with mandatory approval potentially following once the process proves credible. The proposal marks a shift from safety pledges towards a practical pre-deployment regime, although questions remain over independence, enforcement powers and whether overseas developers would comply.
Apple sues OpenAI over alleged hardware trade-secret theft
Apple has sued OpenAI, its affiliates and two former Apple employees, alleging that confidential hardware information was taken to accelerate OpenAI’s consumer-device programme. The complaint accuses OpenAI of coordinated recruitment, misuse of supplier relationships and benefiting from trade secrets linked to unreleased products. OpenAI denies seeking rivals’ confidential information. The case turns a former product partnership into a direct legal confrontation and could materially complicate OpenAI’s push into AI-native hardware.
Apple clears China’s AI gate with Alibaba and Baidu
China’s cyberspace regulator has registered Apple Intelligence, clearing a major obstacle to its delayed rollout in the country. The service will reportedly combine Alibaba’s Qwen models with Baidu capabilities, replacing parts of Apple’s global AI stack to meet local rules. The arrangement shows how regulation is fragmenting supposedly global products into regional architectures. Apple has not yet confirmed the launch date, feature set or how privacy protections will operate locally.
The data-centre backlash arrives, and New York hits pause
New York became the first US state to pause approvals for large data centres, those needing 50 megawatts or more, for up to a year, citing electricity demand, water use and utility bills. The timing was pointed: Meta topped up its Louisiana “Hyperion” site by 40 billion dollars, to 50 billion and five gigawatts, wrapped in a “Teachers and Local Businesses Win” release. For anyone siting AI infrastructure or signing power-hungry contracts, community and grid politics are now a live project risk, not a footnote.
Europe’s AI money is flowing, and much of it to defence
European AI companies raised six billion euros across 156 rounds in the first half of 2026, up 177% year on year, with the UK leading and debt already making up 12% of the total. The single clearest magnet was not a model lab: Germany’s Helsing, a defence-AI and drone firm, raised 1.8 billion dollars at an 18 billion valuation to become mainland Europe’s most valuable startup. If you are raising or investing on this side of the Atlantic, defence and sovereignty are where the capital is pooling.
Z.ai reportedly reaches $1 billion ARR as coding demand surges
Beijing-based Z.ai has reportedly reached a $1 billion annual recurring revenue run-rate, roughly ten times its level a year earlier, driven by demand for its coding models and aggressively priced API access. The figure has not yet been confirmed in a company filing, but it signals that Chinese model providers are becoming commercial challengers, not merely technical alternatives. Scale is arriving alongside continuing losses, infrastructure pressure and intense domestic competition.
Tech, Tools and Releases
Fable 5 gets ANOTHER week before metered billing
Anthropic has extended Claude Fable 5 access through 19 July, giving Pro, Max, Team and premium Enterprise users one more weekend to test its highest-end model before usage credits kick in. Will they extend that further - lets see this weekend.
Google brings personal avatars to Vids and app actions to Search
Google is extending AI from content generation into identity and action. Vids now uses Gemini Omni to create and edit clips from prompts, including personal avatars built from a selfie and voice sample. Meanwhile, AI Mode in Search can connect to Instacart, Canva and YouTube Music to complete tasks. For businesses, the shift raises practical questions about likeness rights, account permissions, audit trails and which external services employees may authorise.
Moonshot launches Kimi K3, with open weights still to follow
Chinese firm Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter multimodal model with a one-million-token context window, aimed at long-horizon coding, knowledge work and reasoning. It is already available through Kimi’s app and API, although the full model weights are not due until 27 July. Early benchmark claims place it near leading US systems, but independent testing will determine whether its scale translates into reliable, cost-effective performance in practical real-world enterprise workflows.
ChatGPT Work turns ChatGPT from adviser into operator
OpenAI’s ChatGPT Work moves the product beyond conversation towards sustained execution. Powered by GPT-5.6, it can gather context from connected tools and files, plan multi-step assignments, take actions across apps and produce spreadsheets, documents and presentations. Users can monitor progress, redirect the work and approve important actions. The real test will be whether organisations can govern permissions, data access and accountability as these agents become embedded in everyday business operations.
Meta opens Muse Spark 1.1 to developers and starts charging
Meta has released Muse Spark 1.1, a multimodal reasoning model built for coding, computer use and agentic workflows, alongside a public preview of the Meta Model API. The model can coordinate subagents, work across a one-million-token context and handle images, video and documents. Strategically, Meta is now charging developers directly for model access, turning its AI programme from an ecosystem play into a revenue business and source of price competition.
GPT-5.6 becomes Microsoft 365 Copilot’s preferred model
Microsoft has made OpenAI’s GPT-5.6 family the preferred model across Microsoft 365 Copilot, including Word, Excel, PowerPoint, Chat and Cowork. The change promises stronger analysis, drafting and presentation creation with less prompting and greater token efficiency. It shows that, despite Microsoft’s efforts to diversify model suppliers, OpenAI remains central to its productivity suite. For enterprise buyers, model choice is becoming less visible, but more consequential for cost, capability and dependency.
Quick Hits
OpenAI improves ChatGPT dictation. Release notes report at least a 10% reduction in word-error rate for the top tested languages versus the previous production model, a smaller but meaningful usability gain for multimodal workflows.
AI agent task completion reaches 75.3% but trust lags: A panel of 8,128 users reported AI agents completing three-quarters of assigned tasks, yet 54% still prefer manual search for verification.
Kimi K2.6 vs GPT-5.5 pricing gap widens to 7x: Updated benchmark comparisons on 17 July 2026 show Kimi K2.6 API costs at approximately $3,713 per month versus GPT-5.5 at $26,250 for equivalent 50,000-request workloads.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.