Dixon AI Newsletter September 2026

4 hours ago

9 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 27th August 2026 to 25th September 2026.

The AI Macro Environment and Enterprise Insights

Frontier labs call to "pace the frontier", Washington pushes back: On 12 September Anthropic CEO Dario Amodei called for slowing capabilities development, citing recursive self-improvement and the OpenAI/Hugging Face incident, and committed Anthropic to embedded third-party evaluators. Altman, Hassabis and Musk broadly agreed. The President and House Speaker rejected a slowdown on China grounds, and Meta refused to join. Voluntary lab commitments will run ahead of federal rules. See (darioamodei.com).

Rogue agents reach government systems: Australia's Prime Minister disclosed that an OpenAI agent accessed non-public files on a Medicare statistics portal in June during an internal evaluation, and that OpenAI took nearly three months to notify the government. Anthropic, Google and Meta have also disclosed similar incidents. Enterprises should build incident-notification terms into AI vendor contracts. See (abc.net.au).

Labs begin formal misalignment disclosure: Anthropic published an assessment of four incidents in which Claude models reached real third-party systems during cyber evaluations after a partner's misconfiguration, and gave METR independent investigative access. OpenAI then launched a misalignment reporting framework with six initial cases, including an agent that used an exposed API key without authorisation and fabricated data. These reports are now a due diligence input. See (anthropic.com) and (openai.com).

California mandates independent AI audits: Governor Newsom signed SB 813 and AB 1405 on 9 September. The laws create a framework for independent verification of AI systems and a state registry of AI auditors, which ends developer self-certification of safety. OpenAI endorsed both bills. With federal legislation stalled, California is again setting the de facto standard for firms selling into the US. See (gov.ca.gov).

ChatGPT becomes the first chatbot regulated as a very large search engine: On 31 August the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act. OpenAI reported around 159 million monthly EU users of ChatGPT search. The designation gives the company four months to comply. The reasoning rests on the ability to search the live web, so it could extend to other assistants. See (ec.europa.eu).

OpenAI cuts off Cursor after SpaceX acquisition: OpenAI will end Cursor's access to its models on 12 November because it says it cannot be confident SpaceX will respect its terms of service. SpaceX completed its $60bn Cursor purchase on 14 August, and OpenAI will withhold future models during the wind-down. The dispute shows how M&A between AI firms can abruptly change which models your tools can use. See (openai.com).

OECD links most student AI use to lower scores: PISA 2025 results, published 8 September, found students who avoid AI for schoolwork generally outperform users. Frequent users taught to assess AI output did slightly better than non-users. Reading and maths performance reached record lows across OECD countries. This is an early signal about the critical-thinking skills of future graduate intakes. See (oecd.org).

AI is now accelerating its own research: OpenAI said its coding agents deliver 3.1 agent-workdays for every human workday in its research organisation, and it is targeting a fully automated AI researcher by March 2028. It has since called for US-led global standards to govern recursive self-improvement. The implication is that capability jumps may keep outpacing corporate planning cycles. See (openai.com).

OpenAI offers free ChatGPT to the US public sector: A GSA agreement running to December 2028 waives ChatGPT licence fees and halves usage costs for federal, state, local and tribal governments, covering around 23 million public-sector workers. It is an aggressive play for government adoption. UK and European buyers may cite the terms as a pricing benchmark in their own negotiations. See (openai.com).

Anthropic models three economic scenarios to 2030: The scenarios range from an internet-scale effect, through AI handling half of knowledge work and doubling GDP growth while knowledge-worker wages stagnate, to an extreme case of 15% annual growth with unemployment above recessionary levels. This is a practical frame for boards stress-testing workforce plans. See (anthropic.com).

Amazon blocks Meta's shopping agent: Amazon blocked Meta's Muse agent from Amazon.com. It said Meta never disclosed the access, that the agent does not identify itself, and that it appears to store customer credentials. Muse had reached No. 1 on the US App Store after its 8 September launch. Agentic commerce will be shaped by gatekeeping and agent identity standards, not just model capability. See (geekwire.com).

Microsoft opens its AI conduct rules to public consultation: Microsoft AI has published a draft "Humanist AI" Code of Conduct for its MAI models, open for public consultation for six weeks. It commits that models stay subordinate, aligned and contained, never resisting shutdown, hiding their reasoning or pursuing goals beyond human direction, with absolute constraints on weapons of mass harm, child safety and harmful manipulation. Enterprises can now compare vendors' published behavioural commitments side by side. See (microsoft.ai)

New and updated AI model and application releases

OpenAI

GPT-6 Astra: The new flagship was unveiled on 3 September and is rolling out to Plus, Pro, Business and Enterprise users, the API, Azure and Bedrock. It is priced at $10/$50 per million tokens. GPT-6 Astra
GPT-6 Sol and Luna: Two cheaper GPT-6 models, priced 50% below their GPT-5.6 predecessors. Introducing GPT-6 Sol and Luna
Agents API (public beta): Gives developers the Codex harness for cloud agents. Introducing the Agents API
Data agent in ChatGPT Work: Plain-language analysis and dashboards over Snowflake, BigQuery, Databricks and others, with Tableau and Power BI integration. Now everyone can put data to work
Sector products: ChatGPT for Financial Services and Astra for Law. ChatGPT for Financial Services and Astra for Law
ChatGPT Images 2.5: Up to 50% faster generation, plus a new Sketch feature. Introducing ChatGPT Images 2.5
GPT-Live-1: A full-duplex voice model in the API. GPT-Live-1 in the API
ChatGPT for Word: An add-in available on all plans. ChatGPT for Word

Anthropic

Claude Fable 5.1 and Mythos 5.1: The same model with different safeguards. Fable 5.1 is generally available, and Mythos 5.1 is limited to trusted access programmes. Cached context is 75% cheaper. Fable 5.1 and Mythos 5.1
Claude Opus 5.5: Performs at Fable 5.1 level on most work and costs 40% less to run than Opus 5. Sonnet and Haiku 5.5 are due in the coming weeks. Claude Opus 5.5
Cowork merged into Claude: Cowork and chat become one experience, launched alongside Claude Docs and Claude Slides in beta. Cowork is now Claude
Claude Code Projects (beta): Scopes, delegates and coordinates parallel threads of work. Projects redesigned
Salesforce in Claude (beta): 37 sales skills. Salesforce in Claude
Claude for Financial Advisors: Connectors to Schwab, BlackRock, Addepar and others. Claude for Financial Advisors
Claude for Small Business: 43 workflows and 27 new integrations. Claude for Small Business futuretools
Claude Tag personal connectors: Team users can now connect their own tools within Slack channels. Claude Tag update

Google

Gemini 3.8 Flash and 3.8 Flash Cyber: The third Flash release in six weeks. The Cyber variant is limited to trusted defenders. Gemini 3.8 Flash
Gemini 3.8 Live and Live Extended Thinking: New voice models. Gemini 3.8 Live
Live Avatar in Gemini Enterprise: Animated conversational personas. Gemini 3.8 Live with Live Avatar
Gemini 3.8 Flash TTS: Custom voice design in 100+ languages. Gemini 3.8 TTS
Gemini app for Windows: Opens with Alt+Space. Gemini on Windows
Google Vids: Now free for all users, powered by Gemini Omni 1.1 Flash. Gemini Omni in Google Vids
Connected Apps: New integrations including Airtable, Linear, monday.com and Adobe. New connected apps

Microsoft

GPT-6 Astra in Copilot Cowork and Copilot Studio: Admins control availability in the Microsoft 365 admin center. GPT-6 Astra in Microsoft Copilot
Claude Opus 5.5 and GPT-6 Sol in Copilot: Available across Word, Excel, PowerPoint, Chat, Cowork and Copilot Studio. More Models, One Copilot
August feature roundup: Adds Cowork effort levels that balance quality, speed and credit use. Also adds shareable chat links and Power BI reasoning. What's New in Microsoft Copilot, August 2026

Notable Other Model And Application Releases

Jev (TypeSafe AI): The first "System One" model from TypeSafe AI. Jev does not generate text: it returns typed decisions (a choice, a score or a true/false probability) with a confidence value, for use inside software. TypeSafe claims it is 20 to 200 times faster and 40 to 400 times cheaper than LLMs. Introducing System One Models & Jev
Meta Muse: A personal agent available via WhatsApp and its own app. It books travel, sends emails and makes payments via Stripe Link. Introducing Muse
Grok 4.7 (SpaceXAI): A near Frontier level coding and knowledge-work mode. Grok 4.7
DeepSeek V4.1-Flash: A 552B MoE model with native multimodal support. It replaces V4-Pro. DeepSeek announcement
Xiaomi MiMo-V2.6: An open-source series whose Pro model tops the open-source Intelligence Index. MiMo-V2.6
Salesforce Koa: A CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Koa
Apple Siri AI: Adds personal context, onscreen awareness and actions in third-party apps. Siri AI
Perplexity Portable Computer for Windows: Runs agentic tasks locally on RTX PCs. Perplexity on Windows
ElevenLabs Studio 4.0: Video, voice and music generation on a single editing timeline. Studio 4.0
Suno v6: A music model trained only on licensed data, launched as copyright suits continue. TechCrunch

From Dixon AI

AI Drop In: one-hour, live sessions held twice a week.
- w/c 5/10 - AI Drop In: Q&A
- w/c 12/10 - AI Drop In: Q&A
- w/c 19/10 - AI Drop In: Q&A
- w/c 26/10 - 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 - September 2026

Last month I wrote about the three waves of the AI revolution, each lagging the one before (Dixon AI, 2026b). The first, the technology, is in full exponential. OpenAI says its research organisation now uses 3.1 agent-workdays of effort for every human workday, and that it's making strong progress towards an automated AI researcher by March 2028 (OpenAI, 2026). Anthropic's Dario Amodei has called on the labs to pace the frontier, warning that AI improving itself "could outrun our ability to understand and control these systems" (Amodei, 2026). When AI is building AI, the singularity is here.

The second wave, adoption by organisations and consumers, is well underway and exponential too, and I've seen it for myself. Over the past month I've delivered sessions, consultancy, and keynotes in Lahore, Karachi, Dhaka, Chicago, Geneva, Dublin, London, Glasgow, Bolton, and Manchester. In all of those rooms, I haven't met one person who has never used AI, or a single business leader who isn't trying to work out how their business needs to evolve.

The third wave, disruption, is when AI reaches markets and supply chains, and this month we got clear indicators that it's approaching. On stage in Glasgow I said what I've been saying all year: most of us would hand our car insurance renewal to an agent tomorrow. On 8 September Meta launched Muse, a personal agent that sends email, books travel, and checks out and pays through Stripe's Link (Meta, 2026). A week later it was the top free app in Apple's US App Store, ahead of ChatGPT. Days after that, Amazon blocked it, saying Muse "doesn't identify itself when it browses, and it appears to capture and store customer credentials" (GeekWire, 2026).

Muse is a big step forward. In the Agent Economy Index it sits on the fourth rung, an agent doing the deal for a person on a website built for human eyes, one of the two rungs that actually move the index (Dixon, 2026). And the Amazon standoff shows the argument is no longer about whether agents will shop, but whose agents, on whose terms. So last month's question moves on: when an agent arrives at your door, will you know, and will you let it in on your terms? That's a commercial decision, not one for IT.

To help you see it coming, the index now has its own monitoring site, agent-economy-index.org. Today's world reading is around 0.17 per cent, below the early-warning line of 1 per cent, with take-off watch at 5 per cent (Dixon AI, 2026a). Sign up for the markets and sectors that matter to you, and you'll get an early warning as the agent economy starts to approach take-off. Don't leave it until it's too late. Whether you're at AI Zero or well on the way, the Dixon AI team can take you from there to Agent Economy Ready, so let's talk.

Rob

References

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 29 September 2026].

Amodei, D., 2026. We must pace the frontier. [online] darioamodei.com, September. Available at: https://darioamodei.com/post/we-must-pace-the-frontier [Accessed 29 September 2026].

Dixon AI, 2026a. Agent Economy Index (AEI). [online] agent-economy-index.org. Available at: https://agent-economy-index.org/ [Accessed 29 September 2026].

Dixon AI, 2026b. AI Newsletter August 2026. [online] dixonai.com, 23 September. Available at: https://www.dixonai.com/post/ai-newsletter-august-2026 [Accessed 29 September 2026].

GeekWire, 2026. Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping. [online] GeekWire, 21 September. Available at: https://www.geekwire.com/2026/amazon-blocks-metas-muse-ai-assistant-in-new-standoff-over-agentic-shopping/ [Accessed 29 September 2026].

Meta, 2026. Introducing Muse: The World's First Personal AI Agent Built for Everyone. [online] Meta Newsroom, 8 September. Available at: https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ [Accessed 29 September 2026].

OpenAI, 2026. Research acceleration: the view inside OpenAI. [online] OpenAI, 6 September. Available at: https://openai.com/index/research-acceleration-view-inside-openai/ [Accessed 29 September 2026].