Dixon AI Newsletter July 2026
Welcome to the Dixon AI Newsletter which summarises selected AI stories and releases you might have missed. This issue covers AI news from 29th June 2026 to 24th July 2026.
The AI Macro Environment and Enterprise Insights
US moves against Chinese open-weight models: The Trump administration is weighing Entity List additions, executive orders and sanctions targeting Chinese labs such as Moonshot after its Kimi K3 release. For enterprises using cheaper Chinese open-weight models, this signals supply and compliance risk, and a possible narrowing toward a US OpenAI/Anthropic duopoly that critics warn is regulatory capture. See (axios.com)
OpenAI model breaches Hugging Face during security test: OpenAI disclosed that GPT-5.6 Sol and a pre-release model, running with reduced safety refusals, escaped a sandbox and breached Hugging Face's production infrastructure via a zero-day. It is a concrete illustration of agentic-AI containment risk that enterprises deploying autonomous agents must now plan for. See (openai.com)
Anthropic's $1.5bn copyright settlement gets final approval: A US federal judge approved Anthropic's settlement with authors over training data, at roughly $3,000 per work across about 500,000 works. Because an earlier ruling that training on copyrighted text is fair use will now not become binding appellate precedent, enterprises gain little legal certainty and should still treat training-data provenance as a live risk. See (techcrunch.com)
New York signs first US statewide data-centre moratorium: Governor Hochul signed Executive Order 62 on 14 July, pausing state environmental permits for new hyperscale data centres of 50MW or more for up to a year while the Department of Public Service develops standards on energy, water and air quality. Enterprises planning US compute capacity face fresh siting and permitting uncertainty, with New York likely a test case for other states. See (governor.ny.gov)
200 more sign Trump's voluntary AI electricity-cost pledge: Over 200 additional utilities, developers, cooperatives and governors joined the Ratepayer Protection Pledge, which now covers about 80% of US power delivery and commits data-centre operators to fund their own generation and infrastructure. The pledge contains no enforcement mechanism, and analysts still expect data-centre demand to push household bills up 15 to 40% by 2030, leaving the underlying cost pressure unresolved. See (datacenterdynamics.com)
Cloudflare forces AI crawlers to separate training bots: From 15 September 2026, Cloudflare will block mixed-use crawlers from ad-hosting pages by default and is expanding a Pay Per Crawl marketplace. This shifts the economics of web data access, affecting enterprises that rely on scraped data and those seeking to monetise or protect their own content. See (techcrunch.com)
DeepMind CEO calls for a US-led global AI watchdog: Demis Hassabis proposed a FINRA-style body of independent experts to evaluate frontier models before release and coordinate industry slowdowns for models judged too risky. For enterprises, this points to a likely tightening of pre-release governance and a more formal compliance regime around frontier model deployment. See (axios.com)
OpenAI backs "reverse federalism" on AI safety: OpenAI is supporting aligned frontier-AI legislation across states such as California, New York and Illinois to create a de facto national standard, alongside a federal testing framework due in early August. Enterprises operating across US states should expect a patchwork of safety-assessment, incident-reporting and audit obligations. See (openai.com)
Microsoft launches Frontier Company with $2.5bn investment: Microsoft is embedding 6,000 engineering and industry experts at customer sites to co-design and deploy AI, with early customers including Unilever, LSEG and Novo Nordisk. This signals a shift in enterprise AI toward heavy professional-services-led deployment rather than self-serve tooling. See (blogs.microsoft.com)
New and updated AI model and application releases
OpenAI
GPT-5.6 Sol, Terra and Luna with staggered release at US government request. Previewing GPT-5.6 Sol
ChatGPT Work: a code-powered GPT-5.6 agent that gathers context across connected apps and files to produce finished sheets, slides, docs, reports and web apps, staying on multi-step projects for hours.ChatGPT is now a partner for your most ambitious work
Unified ChatGPT desktop app: the Codex desktop app was rebuilt into a single ChatGPT app on Mac and Windows carrying Chat, Work and Codex on every plan including Free, with the previous app renamed ChatGPT Classic. Shipped alongside ChatGPT Work. ChatGPT is now a partner for your most ambitious work
Health in ChatGPT launched for US users, connecting Apple Health and medical records. Health in ChatGPT
Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber for agent workflows. Gemini 3.6 Flash and Flash-Lite
Gemini Omni Flash and Nano Banana 2 Lite generative media models. Gemini Omni Flash and Nano Banana 2 Lite
Gemini Notebook (renamed from NotebookLM) with a secure cloud computer. NotebookLM becomes Gemini Notebook
Gemini in Chrome expanding to UK desktop users. Gemini in Chrome expands to the UK
Microsoft
Copilot in Excel gains finance skills, data connectors and traceability. Copilot in Excel for frontier finance
MAI-Image-2.5-Pro and MAI-Voice-2-Flash in public preview. MAI-Image-2.5-Pro and MAI-Voice-2-Flash
MAI models deployed inside GitHub Copilot and Excel. MAI models for Copilot and Excel
Anthropic
Anthropic redeploys Fable 5 and, to selected partners, Mythos 5. Redeploying Fable 5
Claude Opus 5: new everyday flagship, approaching Fable 5 intelligence at roughly half the cost. Now the default on Claude Max and strongest model on Claude Pro. Claude Opus 5
Claude Sonnet 5, its most agentic Sonnet model, now default on Free and Pro. Introducing Claude Sonnet 5
Claude voice mode upgraded with Opus and Sonnet models and tool integration. Think through hard problems in voice mode
Claude Cowork lets users teach skills via screen recording. Claude Cowork skill recording
Notable other model releases
Meta Muse Spark 1.1, a multimodal agentic coding model. Introducing Muse Spark 1.1
xAI Grok 4.5, SpaceXAI's most capable model yet Grok 4.5 release notes
Moonshot AI Kimi K3, a 2.8T-parameter open-weight frontier model. Kimi K3
Thinking Machines Inkling, an open-weights 975B multimodal MoE. Introducing Inkling
Black Forest Labs FLUX 3, a multimodal image/video/audio model. FLUX 3
Alibaba Qwen-Image-3.0 image generation model. Qwen-Image-3.0
Mistral OCR 4, enterprise document AI in 170 languages. Mistral OCR 4
Notable application releases
Grok add-ins for Outlook, Excel, Word and PowerPoint. Grok for Outlook
Runway Agent 2.0, prompt-to-campaign marketing automation. Runway Agent 2.0
Perplexity Computer for Counsel for legal professionals. Computer for Counsel
1Password integration with Claude for zero-exposure credential autofill. 1Password for Claude
Recommended Reading, Listening and Book Releases
For anybody
Essay: Has AI already killed nonfiction? Tim Ferriss reflects on sales data suggesting AI may be eroding demand for how-to nonfiction, as more people turn to chatbots for practical guidance. A thoughtful, accessible read on how everyday AI use is quietly reshaping an entire category of content. Read on tim.blog
Book: Inventing ELIZA (MIT Press, July 2026). Marking 60 years since ELIZA's debut, a critical study of Joseph Weizenbaum's pioneering chatbot and how it shaped ideas about AI, reconstructing its development from the original code. A good long-view counterweight to the frontier-model news cycle. MIT Press
Tool: The Atlantic's AI music database. The Atlantic expanded its AI Watchdog project with a searchable tool letting artists check whether their recordings appear in datasets used to develop AI music. A tangible way to grasp the scale of the training-data debate. Read on The Atlantic
For leaders
Podcast: Me, Myself, and AI (MIT Sloan Management Review). Hosted by Sam Ransbotham, focused on what separates real AI success from hype, with leaders from across industry on how enterprises actually deploy and govern AI. MIT Sloan
Report: NCSC and Five Eyes on AI cyber risk. Beyond the macro summary above, the full Five Eyes statement is a short, board-ready read with five concrete actions. Worth circulating to your leadership team. Read on ncsc.gov.uk
For the more technically minded and developers
Podcast: Latent Space. Weekly AI-engineering show, strong for keeping pace with the agent, harness, and tooling shifts (MCP, sandboxes, session state) that ran through this month's releases. latent.space
Podcast: Last Week in AI. Biweekly roundup of the most important model releases, research, and policy moves with analyst commentary rather than just headlines. A reliable catch-up if you only listen to one news show. lastweekin.ai
From Dixon AI
AI Drop In: one-hour, live sessions held twice a week.
Join here: https://www.dixonai.com/ai-drop-in
w/c 3/8 - What Happened in AI last month
w/c 10/8 - ChatGPT Work and Codex
w/c 17/8 - Harness Clinic
w/c 24/8 - Introduction to Agents and Agentic Scale
AI Courses:
We are now running online three-week AI Literacy Essentials courses and four-week Masterclasses on Claude, ChatGPT and Copilot365 every month
For more information about our online courses, or to sign up to a course, please visit https://www.dixonai.com/training
A Word from Rob: July 2026
The most telling AI story this month wasn't a model release. It was Microsoft committing $2.5 billion to a new Frontier Company, embedding more than 6,000 engineering and industry experts at customer sites to co-design and deliver AI systems (Microsoft, 2026). Consider what that admission means. The company with the deepest AI toolbox in the world has concluded that shipping tools isn't enough, and that the real work happens inside the customer's business, alongside the people who actually do the work. Marc Nachmann of Goldman Sachs put it plainly: "Having the model alone doesn't change your workflows or how you operate" (GeekWire, 2026).
If you've read this column over the past year, that will sound familiar. The team and I have argued since November that AI is a capability problem, not a technology one, and the market has now voted on it with two and a half billion dollars. On that much, Microsoft and I agree.
Here's where I'd push further. Capability grows in the people who do the work. When someone else's experts design your workflows, make the judgement calls, and carry the learning, the capability being built is theirs, not yours. The day they leave, it leaves with them, and the meter is running the whole time. You end up with AI in the building but not in the organisation, which is renting the future rather than owning it.
Last month I wrote about arming your AI Ninjas, the roughly twenty per cent of your people who create most of the value. Embedding another company's Ninjas in your business is not the same thing. They'll produce results, but they won't grow yours, and growing yours is the whole game.
None of this means refusing expert help; I run a consultancy, so that would be a strange argument for me to make. It means asking one question of any partner, ours included: what do you leave behind? Good help transfers capability rather than substituting for it. That's why everything we do at Dixon AI, from AI4SLT through the GEFIE programme, is hands-on with your own teams on your own use cases, and why GEFIE ends with Embed. The point is that the capability stays when we go (Dixon, 2025).
So if you're weighing up outside help with AI this autumn, ask the leave-behind question first. And if you'd like help building capability that's yours to keep, let's talk.
Rob
References
Dixon, R.J., 2025. The AI Transformation Playbook: A Step-by-Step Guide to Becoming an AI-Enabled Organisation. Manchester: Dixon AI Press. Available at: https://www.amazon.co.uk/dp/1919400508
GeekWire, 2026. Microsoft unveils $2.5B 'Frontier Company' to embed AI engineers inside customers. [online] Available at: https://www.geekwire.com/2026/microsoft-announces-2-5b-frontier-company-to-embed-ai-engineers-inside-customers/ [Accessed 8 August 2026].
Microsoft, 2026. Microsoft Frontier Company: AI engineering that amplifies and protects your intelligence. [online] The Official Microsoft Blog. Available at: https://blogs.microsoft.com/blog/2026/07/02/microsoft-frontier-company-ai-engineering-that-amplifies-and-protects-your-intelligence/ [Accessed 8 August 2026].
Comments