In partnership with
Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
A huge week in AI with the release of two frontier models. OpenAI's GPT-6 Astra arrived days after Claude Fable 5.1 launched.
These models will keep leapfrogging each other, and every release sends someone to your desk asking for another licence. Chasing whichever model is winning this week is a losing game. The advantage that lasts is the context, skills, data and workflows sitting where your people already work, so you can move between models as needed. Not everyone needs the best-of-breed model for every job.
It was impressive to see the creative uses of OpenAI’s Astra model. Enjoyed this one in particular:
1:1 AI Coaching Programme: Space opening in my next four-week programme. Built around your AI stack, from ChatGPT and Claude through to Claude Code, Codex and Copilot Studio, and your actual workflows rather than a generic curriculum. Reply if you want the details.
Happy reading ✌️
Did someone forward you this? Sign up.
🇳🇿 New Zealand News
Wellington mayor Andrew Little says “large chunks” of a $435,000 Deloitte report on council staffing were written by AI. Deloitte double-counted employees against vacant roles, used three-year-old Taxpayers’ Union data, counted part-time roles at full-time salary and overstated staffing expenses by $21.5 million.
3 min read
Our take: This says little about the technology and a lot about the process. Australia saw the same behaviour, so it is a trend in how consulting firms work, not a fault in the tool. The step that vanished is the one where a person checks the cited thing exists, and clients still pay for it in the fee. Deloitte Australia refunded A$97,000 on a A$440,000 contract for the same failure, which barely moves the incentive to check.
New Zealanders send almost 10 million ChatGPT messages a day, and the top 10% of firms produce 14 times more output per worker. OpenAI’s NZ data shows 40% of surveyed users on it daily, and nearly half of enterprise output coming from agentic products. Gawesha Weeratunga says the gap between leading and lagging firms is widening.
4 min read
Our take: Growth is a good sign, but NZ trust in AI surveys stays low, so these figures probably lean on power users. They are also vanity metrics. OpenAI’s own August report found revenue per employee has no meaningful link to tokens or messages, and no CFO signs off spend on a message count.
Squirrel chief executive David Cunningham received more than 1,000 applications for a single role, most of them clearly AI-generated. He says cover letters are now “a waste of time really”.
4 min read
Our take: Getting a job is competitive and getting more so. When supply grows, the filtering process has to change with it. Clearly the intake process is not keeping pace. It also pushes both sides back to meeting face to face, another case of AI raising the value of the human aspects of a process.
An independent commissioner has decided Southland District Council will not review the consent conditions for Datagrid’s AI data centre at Makarewa. The Southland Sustainable Resource Coalition applied under section 128 of the Resource Management Act over low-frequency noise, infrasound and cooling noise. The commissioner found the information did not meet the statutory threshold.
3 min read
Fonterra’s Global Ingredients division retired a decade-old analytics system and moved to Databricks across more than 100 export markets. The old system was consuming about half the team’s engineering time on maintenance. The division put Databricks Genie agents over the data and delivered milk supply and pricing forecast analysis in the first month.
4 min read
Our take: The agents were built after the underlying platform work was completed. That order is why Fonterra shipped the use case fast. Reality is most Kiwi businesses don’t have the resources, capability and runway to build a solid foundation before executing on AI. These case studies are interesting but not reflective of the worlds of 99% of NZ businesses.
The Reserve Bank named Anthropic’s frontier model Mythos in its Financial Stability Report, the first time it has singled out a specific AI model. It flagged frontier models amplifying cyber threats and the concentration risk of NZ’s financial system depending on a few overseas AI providers. FSB chair Andrew Bailey later told G20 ministers frontier AI cyber risk is the most immediate threat to the global financial system.
4 min read
Our take: The concentration risk is the part that applies to every business here, not just banks. Most of the country runs on three overseas providers.
Auckland startup Thread launched a marketing system of record, with DB Breweries, FreshChoice, ANZCO Foods and Metlifecare already on it. It stores campaign briefs, media plans, creative and performance reports in one place, queried through a conversational layer. Rather than pointing a general model at unstructured files, it queries structured marketing data.
3 min read
Our take: Thread’s line about plugging AI into an unorganised server is the whole problem in one sentence. Structure the data first and the model has something worthwhile reading from to deliver insightful responses from. A critical misstep in most people’s AI use.
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
📚️ Mike’s Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ Implementing AI is like skiing: easy to start and hard to master: Ramp says the top 1% of businesses account for roughly 80% of OpenAI and Anthropic spend. The bigger insight is how much work still sits between buying licences and redesigning business workflows.
Read
2️⃣ GPT-6 Astra makes model choice a losing game: GPT-6 Astra landed days after Claude Fable 5.1, giving employees another reason to demand the latest licence. The durable advantage is portable context, data and workflows that can move between models.
Watch
3️⃣ Frontier intelligence gets cheaper while implementation gets harder: Claude Fable 5.1 cuts cache costs and delivers near frontier performance at lower effort. That matters for agents, but every firm can buy the same model; the real expense now sits around implementation.
Read
4️⃣ I’ve built my own AI coach for Coast to Coast: I’m using 26 years of Strava data to build a personal training app that combines Garmin activity data with Google Calendar, updates the plan as I train, and schedules each session around real life.
Read
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Claude Fable 5.1 launched generally available, with Mythos 5.1 restricted to vetted professionals through the Cyber Verification and Life Sciences Verification programmes. (Sep 1)
Pricing is $10 per million input tokens and $50 per million output, with cache reads cut 75% to $0.25 per million. (Sep 1)
Commerce Agents shipped as Apache-2.0 blueprints for shopping and merchant agents, with Shopify, Priceline, Visa and Mastercard integrating. (Sep 2)
NYSE told the House Financial Services Committee it used Anthropic’s Mythos model to find cyber flaws, some sitting in the codebase for decades. (Sep 2)
Anthropic plans to file its IPO prospectus after Labor Day, targeting a listing as soon as late September. (Sep 3)
OpenAI
GPT-6 Astra launched as the new flagship, with state-of-the-art results on computer use, software engineering, cybersecurity and science. (Sep 3)
Astra is the first model to meet OpenAI’s critical cybersecurity capability threshold, so the public version refuses certain security prompts. (Sep 3)
Astra scored 99.9% on ARC-AGI-3 with enhanced tools against 7.8% for GPT-5.6 Sol, and 100% on ExploitBench against 78.5%. (Sep 3)
OpenAI committed $1 billion in subsidised Daybreak access, training and support for water utilities, grid operators, local government and community banks. (Sep 3)
Zendesk and OneNote plugins entered beta for Business, Enterprise and Edu, letting teams handle tickets and notes inside ChatGPT. (Sep 3)
Gemini 3.8 Flash launched at the same introductory price as 3.7 Flash, $0.75 per million input tokens and $3.75 per million output. (Sep 2)
Gemini 3.8 Flash Cyber shipped as Google’s most capable cybersecurity model, writing 2.6 times more correct Chrome patches than the best commercial models. (Sep 2)
Agentic video understanding lets Gemini navigate video timelines on demand, cutting analysis costs up to 66% and token use up to 88%. (Sep 1)
Google Vids now turns Docs, PDFs and Word files into short videos, generating script, narration and visuals automatically. (Sep 2)
A federal court rejected the Justice Department’s proposed break-up of Google’s ad tech business, refusing the forced sale of AdX. (Sep 2)
Chinese Frontier Labs
Alibaba released Qwen3.8-Max-0902, lifting its CodeArena front-end score 22 points to 1,691 and taking first place on that leaderboard. (Sep 2)
Moonshot AI filed confidentially for a Hong Kong IPO, seeking roughly $3 billion at about a $50 billion valuation. (Sep 3)
Alibaba released Qwen-Drive-1.0-4B under Apache 2.0, a driving foundation model scoring 77.8% on LingoQA and 91.4 PDMS on NAVSIM. (Sep 1)
DeepSeek shipped nothing in the window, with V5 still rumour and no changelog entry, API model string or technical report. (Sep 6)
Open-Weight Models
RedMonk found Chinese open-weight models took 63% of OpenRouter enterprise tokens in August, up from 4.5% in early 2025. (Sep 3)
Ollama’s chief executive said open models now carry 80% to 90% of enterprise token volume on 10% to 20% of budget, citing AT&T at 40%. (Sep 4)
Almost no open-weight models launched in the window, with every major release from Sep 1 to Sep 4 proprietary. (Sep 6)
NVIDIA
NVIDIA confirmed it will acquire Hugging Face for $12,930,300,000, a platform carrying over 3 million models and 500,000 datasets. (Sep 3)
NVIDIA and CrowdStrike launched SafeMind, an agentic security system pairing Falcon IQ’s 50-plus agents with Nemotron foundation models. (Sep 1)
NVIDIA invested $3.5 billion in MediaTek convertible bonds, adding it to the NVLink Fusion ecosystem for rack-scale infrastructure. (Aug 31)
Microsoft
GPT-6 Astra shipped day-one across Microsoft Copilot, Copilot Cowork, Copilot Studio, GitHub Copilot and Microsoft Foundry. (Sep 4)
Claude Fable 5.1 reached Copilot Cowork and Copilot Studio with enterprise data protection, plus GitHub Copilot general availability. (Sep 1)
GitHub Copilot code review gained authority to approve pull requests, not just comment on them. (Sep 1)
Admins can now set any available model as the enterprise default, and cap Copilot spend per user with automatic resets. (Sep 2)
Amazon, Apple and Meta
Claude Fable 5.1 became available on Amazon Bedrock and the Claude Platform on AWS, marketed for multi-hour autonomous multi-app work. (Sep 1)
Bedrock AgentCore Identity added a hosted Consent Portal where users approve what resources an agent may access before it proceeds. (Sep 6)
Apple’s Hybrid Compute launched on Mac, splitting a task between cloud models for reasoning and a local model for any step touching private data. (Sep 1)
John Ternus became Apple chief executive, with Tim Cook moving to executive chairman after a last day on 31 August. (Sep 1)
Meta launched Muse Voice Transcribe, its first real-time audio model, covering 70 or more languages at $3 per 1,000 audio minutes. (Sep 1)
✨A few people have asked…
It’s Mike here, I run The AI Corner.
I'm not just into writing about AI. I run Allexive. We work with businesses that have spent on AI licences and training and still can't point at what it changed. We redesign the work so it does.
👋 Mike & Erin