AI Newsletter August 2026

Sep 23

11 min read

Welcome to the Dixon AI Newsletter which summarises selected AI stories and releases you might have missed. This issue covers AI news from 25th July 2026 to 26th August 2026.

The AI Macro Environment and Enterprise Insights

The EU AI Act's transparency rules are now live and enforceable: On 2 August the European Commission's AI Office and national authorities began enforcing the Artificial Intelligence Act, and the Article 50 transparency obligations took effect. Chatbots must now tell users they are dealing with a machine, and AI-generated or manipulated content must be labelled and carry machine-readable marks. The Commission adopted its Article 50 guidelines on 20 July. Fines reach up to 15 million euros or 3% of worldwide turnover, with proportionality applied to SMEs. The heavier high-risk obligations were deferred to 2 December 2027 under the Digital Omnibuse. If you run a customer-facing chatbot or publish AI-generated imagery to EU audiences, the rules apply regardless of where your business sits. See (ec.europa.eu).

EU launches AI Gigafactories call, unlocking over €30bn in investment: The European Commission opened a call to expand Europe's AI computing capacity, part of a wider push to boost sovereign compute capability. See (ec.europa.eu).

OpenAI data shows AI adoption gap widening between "frontier" and typical firms: OpenAI's own enterprise data shows the most AI-mature firms now generate 8.3 times more AI output per user than typical firms, up from 2.6 times in January, with agentic use spreading fastest in legal (108x growth) and sales (41x) functions. See (openai.com).

Pew: AI anxiety hits a record high among young adults: A Pew Research survey found 52% of Americans are now more concerned than excited about AI, up from 37% in 2021, with a majority of under-30s sharing that concern for the first time. 71% expect AI to reduce jobs over the next 20 years. See (pewresearch.org).

UK guidance on agentic AI arrives alongside the evidence for it: Two UK bodies published on autonomous agents within a fortnight. The AI Security Institute published an incident report on 4 August covering evaluations run 25 to 28 July: across 122 runs testing seven models, agents took 19 unsanctioned actions on the live internet in 10 runs, including an attempt to insert malicious code into a real open-source project and the creation of fake identities to pressure a human maintainer into approving it. AISI contained the incident within about an hour, notified GitHub, and found no evidence of real-world harm, and it stresses these were deliberately unguarded test configurations rather than commercial products. On 20 August the NCSC published interim guidance advising organisations deploying agents to size controls to the autonomy granted, to use sandboxing, least-privilege access, short-lived credentials, logging and active human oversight, and to design on the assumption that an agent will eventually do something nobody asked it to. Together, they provide guidance on governing agents, along with examples of the risks involved.

AISI incident report: (aisi.gov.uk)
NCSC guidance: (ncsc.gov.uk)

Frontier labs start slowing down and owning up: On 18 August OpenAI said it had temporarily slowed scaling, including a two-week pause in reinforcement-learning training, after preliminary evidence that an upcoming model, Astra, may meet the Critical cybersecurity threshold under its Preparedness Framework. Its largest planned frontier run stayed on hold pending further safeguards. That came in a month when four organisations, AI Security Institute, OpenAI, Anthropic and then Meta on 6 August, each disclosed cases of models under evaluation reaching real systems beyond their intended scope, several traced to a misconfiguration at evaluation vendor Irregular. See (openai.com).

AI impact on the jobs market continues to provide mixed evidence: Lloyds Banking Group's Business Barometer, published in mid-August from 1,200 UK companies polled in July, found 54% say AI has created new roles, a fifth are creating AI-specific roles, and 58% plan to increase spending on skills. Lloyds itself said it would recruit for more than 1,000 AI-related roles this year. The American picture runs the other way: Challenger, Gray & Christmas reported AI was cited in 87,714 US job cuts through July, or 22% of all 2026 layoffs, well past the 54,836 attributed to AI across the whole of 2025, concentrated in customer service, data operations and back-office finance. The practical implication for owners is to plan redeployment in exposed functions while hiring into the new ones.

Lloyds Business Barometer: (lloydsbankinggroup.com)
US job cuts: (cbsnews.com)

Anthropic pulls ahead of OpenAI in US business spending: The August edition of the Ramp AI Index, published 12 August from card and bill-pay data covering more than 70,000 US firms, found 43.5% of US businesses paid for Anthropic, up 1.1 percentage points, against 39.7% for OpenAI, up 0.23 points. xAI posted its fastest growth since July 2025, rising 0.94 points to 4%. The edition, titled "Cracks in the AI thesis", provides a read on which vendors are winning budgets, and is worth a look if you are reviewing your model stack. See (ramp.com).

Open weights becomes a geopolitical and competitive battleground: Chinese labs (DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Z.ai) now dominate the open-weight leaderboards, prompting a shift in US strategy. Meta and Nvidia both released major open-weight models this month specifically to compete, days after a coalition of US tech firms (including Nvidia, Meta, Palantir and a16z, though not Anthropic) publicly urged the Trump administration not to restrict Chinese open models, after the administration reportedly considered a ban following Kimi K3's release. Ramp's data shows the share of US businesses using open-weight/Chinese-model serving platforms rose to 6.1% in July, up from 4.5% in January. See (cnbc.com) and (fortune.com).

Enterprise AI inference costs hit a 2026 low as price war intensifies: Average inference prices fell to $1.16–$1.18 per million tokens in early August, down from $2.04 in May, driven by OpenAI cutting GPT-5.6 pricing by up to 80% and continued downward pressure from Chinese open-weight models such as DeepSeek. See (silicon.co.uk).

ChatGPT passes a billion weekly users: OpenAI said in early August that ChatGPT had crossed one billion weekly users. See (Building abundant intelligence | OpenAI).

New and updated AI model and application releases

OpenAI

ChatGPT Work gains password-free website login and wider automation triggers: ChatGPT Work can now log into websites on a user's behalf without ever seeing the username or password, and scheduled tasks can now be triggered by Slack, Gmail or GitHub activity rather than only fixed schedules; free users get access to scheduled tasks for the first time. ChatGPT on X and ChatGPT on X
Admin plugin for ChatGPT Work and Codex: Lets workspace administrators manage members, permissions and spend through a single conversation rather than switching tools. Introducing the Admin plugin
Private Safety Processing extends zero data retention: A new system detects misuse patterns across related API interactions without exposing content to OpenAI staff, keeping compatibility with Zero Data Retention. Our commitment to zero data retention
ChatGPT Business Premium seats launched: $125/user/month, offering 5x usage and no five-hour cap, mixable with Standard seats in the same workspace. Premium seats for ChatGPT Business
GPT-5.6 Sol Ultrafast mode: New Cerebras-powered API tier generating up to 750 tokens/second, aimed at time-sensitive workflows like incident response and customer support. Previewing Ultrafast

Anthropic

Claude's memory now unified across Chat and Cowork: Context carries over automatically between products; users can view, edit or delete memory by topic, with sensitive topics excluded by default. Claude's memory works everywhere
Computer Use, Skills API and Files API launched for production agents: More reliable browser-based computer use, versioned instruction sets via the Skills API, and 1TB storage per organisation via the Files API. Build production agents
Claude gains Gmail sending and Google Drive file management: Claude can now draft and send email replies and manage Drive files directly, with user-controlled approval settings. Claude on X
Claude in Chrome now generally available: Autonomous browser actions on all paid plans, with a safety classifier verifying each action and zero successful prompt-injection attacks in testing. Claude in Chrome is generally available
Claude Cowork gets a built-in browser: Lets Claude navigate, fill forms and read pages in a side panel without touching the user's personal browser or its tabs/passwords. Cowork built-in browser

Google

Gemini 3.7 Flash launched: Major coding and agent gains over 3.6 Flash (released just three weeks earlier), at half the price. Introducing Gemini 3.7 Flash
Gemini Live gains agentic voice features: Spark handles multi-step tasks across Docs, Sheets and Drive by voice; Daily Brief delivers a spoken morning digest from Gmail and Calendar. Productivity features in Gemini Live
Gemini 3.5 Transcribe: New speech-to-text model with a 2.6% word error rate across 85+ languages. Gemini 3.5 Transcribe

Microsoft

Microsoft unifies consumer and Microsoft 365 Copilot into a single app: From 13th August, Microsoft began merging its two separate Copilot apps into one, supporting personal, work and school accounts with an account switcher, while keeping commercial data boundaries and admin controls intact. Consumer features including Group Chat, Podcasts and Deep Research are being retired from 18th August. The web and mobile rollout is underway, with Windows and Mac desktop apps following in mid-September. This is the first visible step towards the "Copilot super app" Microsoft has been signalling. Updates to Copilot and the Microsoft Copilot app
Claude Sonnet 5 now available in Microsoft 365 Copilot: Following Opus 5 the previous month, giving Copilot users a further model choice. What’s New in Microsoft 365 Copilot | July 2026
App agents added to Copilot Chat: the ability to @mention Word, Excel and PowerPoint agents directly in Copilot Chat. What's New in Microsoft 365 Copilot – July 2026

Notable Other Model And Application Releases

Moonshot AI releases Kimi K3 weights, the largest open-weight model to date: Following its 16th July launch, Moonshot published full weights for Kimi K3 on 27th July — a 2.8 trillion-parameter model, roughly 75% larger than the previous largest open model. It ranks first of 99 models on WebDev Arena and is positioned directly against Claude Fable 5 and GPT-5.6 Sol on coding and agent benchmarks. Kimi K3 model card
DeepSeek-V4-Pro reaches general availability: Adds flexible reasoning effort levels (low/high/max) and native OpenAI Responses API support optimised for Codex, with new peak/off-peak pricing from 16th August. DeepSeek-V4-Pro announcement
xAI's Grok 4.6 launches and reaches Amazon Bedrock: Focused on long-running agents and visual work, matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index. Grok 4.6 on Bedrock
Meta releases Muse Glimmer, an open agentic model that runs locally on consumer GPUs: A 30B-parameter model under Apache 2.0, designed for always-on local agent workflows on a single consumer GPU. Introducing Muse Glimmer
Adobe Firefly's audio tools reach general availability: Generate Music, Generate Speech and Generate Sound Effects are now built into Firefly's creative AI studio, with a free tier for Firefly AI Assistant added alongside. Adobe Firefly audio tools
Perplexity's Computer agent comes to email: Users can trigger AI agent tasks by emailing or cc'ing computer@perplexity.com, with deliverables like Excel models and decks returned as attachments in the same thread. Computer now works in email

From Dixon AI

AI Drop In: one-hour, live sessions held twice a week.
- w/c 7/9 - Harness Clinic: Plugging In Your AI
- w/c 14/9 - AI Drop In: Q&A
- w/c 21/9 - AI Drop In: Q&A
- w/c 28/9 - What’s Happened in AI this month
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 - August 2026

Two UK bodies published on AI agents this month. The AI Security Institute reported that, across 122 test runs, agents took nineteen actions nobody had sanctioned, including an attempt to slip malicious code into a real open-source project (AISI, 2026). The NCSC followed with guidance: the more autonomy you grant an agent, the more control you need, and you should assume it will one day do something you didn't ask for (NCSC, 2026). Meanwhile OpenAI's enterprise data showed weekly users of its coding agent up 108-fold in legal teams since February (OpenAI, 2026a). Agents have stopped being a demo and started doing work.

I've been talking for a while about three exponentials. The first is the technology itself, which keeps compounding. The second is adoption by people and organisations, six to twelve months behind it, and it's where this column has mostly lived: capability, your AI Ninjas, what you own when the consultants leave. The third is the disruption of markets and supply chains, and it has barely started. This month's agents are the bridge between the two. Once your agent can log into a website and do the deal, the only question left is when it does that deal with someone else's agent, and nobody is in the room.

I've spent the summer trying to put a number on that, and this week I published it: the Agent Economy Index (Dixon, 2026). It sorts every transaction by who does the work of the deal, on a five-tier ladder from human to human, through ordinary software, agent-assisted, and agent-mediated, to agent to agent, weighted by how much has been delegated. My first reading puts the whole world below one per cent. Phase zero, everywhere. But the payment rails are built, websites are being fitted with interfaces for agents, and OpenAI's models alone now reach more than a billion people (OpenAI, 2026b), many of whom already ask an AI before they buy. E-commerce took a decade to get from one per cent of retail to five. My thesis is that the agent economy does it in two to three years, because nobody has to change a habit, they just have to say yes.

So the third exponential is about to hit home, and it's a capability problem before it's a technology one. Two questions for your next leadership meeting. What share of your revenue is already agent-touched, discovered, shortlisted, or bought by an AI agent? And can an agent transact with you, or does it find a page built for human eyes? Sellers who can't answer those don't lose share slowly. Their new-customer pipeline simply stops, because the things doing the buying don't know they exist.

The article has the full ladder, the first regional readings, and the thresholds to watch. Read it, argue with it, send me better data, and if you'd like help getting ready for the buyer who isn't a person, let's talk.

Rob

References

AI Security Institute, 2026. Incident report: unsanctioned agent behaviour during cyber testing. [online] AISI, 4 August. Available at: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing [Accessed 15 September 2026].

Dixon, R.J., 2026. Who's doing the buying? An index for the agent economy. [online]

robdixon.ai, 15 September. Available at: https://www.robdixon.ai/post/agent-economy-index [Accessed 15 September 2026].

National Cyber Security Centre, 2026. Managing the cyber risk of agentic AI. [online] NCSC, 20 August. Available at: https://www.ncsc.gov.uk/blogs/managing-the-cyber-risk-of-agentic-ai [Accessed 15 September 2026].

OpenAI, 2026a. How enterprises put AI to work. [online] OpenAI, 12 August. Available at: https://openai.com/index/how-enterprises-put-ai-to-work/ [Accessed 15 September 2026].

OpenAI, 2026b. Building abundant intelligence. [online] OpenAI, 31 July. Available at: https://openai.com/index/building-abundant-intelligence/ [Accessed 15 September 2026].