VOL 2, ISSUE 34 · 4 SEPTEMBER 2026
If planning, researching and strategising take up, say, 40% of a task, then 60% is left to execute and iterate. That 60% can take up 100% of your available time if you're doing it manually.
This is the week that argument stopped being theoretical for me. The kids are home, my partner is at work, and the writing happens in the gaps — text the agent, deal a new hand of Go Fish, check what came back.
It's also the week I started calling something vibe sourcing: not asking an agent to build you software, but asking it to go and find the open-source thing that already does the job. Two subscriptions came out of my stack that way.
Read the full piece...
Also on the site this week
Security & Risk
Google signed the letter asking for cyber-capable AI, then gated its cyber model
156 companies called for a surge in cyber defence on 27 August. On 2 September Google shipped its most capable security model to a selected set of trusted defenders.
OpenAI Wants Your Trust Back, But Is Unclear On How
OpenAI paused frontier training and wound down its side projects. The product it is refocusing on asks for your calendar, your finances and your card.
Anthropic says it had one layer of defence where it needed several
The post-incident write-up admits it relied on a single layer where it needed several. The independent METR review is still to come.
An agent's tool list shows what you wired into it
A ChatGPT Work session published its tool and skill inventory as a public website. The category names show which integrations that session had been given.
Perplexity brings hybrid frontier and local AI to your Mac
Hybrid compute runs a classifier on your Mac to decide what reaches the cloud. The admin rules and the record of what left are Enterprise features.
Agents & Tooling
Runway's Solaris generates app interfaces frame by frame, with no code
Runway's first Interface World Model generates a UI frame by frame with no code underneath. Accessibility is on its own list of unsolved problems.
The Industry
Nvidia confirmed the Hugging Face deal and promised its compute stays optional
A week ago it was an unconfirmed report. Jensen Huang has now announced it himself, at $12,930,300,000, and committed that Nvidia compute will not be required to use Hugging Face.
Did Multiverse actually build Europe's leading AI model?
A Spanish company launched a 438B reasoning model as Europe's best. Its API changelog, the benchmark it cites as validation, and a Community Note all point at a Chinese open-weight base.
Most of the sources behind Perplexity's software recommendations sit outside the top 100,000 websites
Trellner put 380 software categories to Perplexity and kept every citation. 59.8% pointed at domains ranked worse than #100,000.
World models could widen the AI divide before they improve ordinary work
Far behind language models in deployment evidence, and their development could concentrate further. Businesses will be offered physics-aware systems before anyone can show they work.
Latest from Second Brain Chronicles
I’m Spending $50 to See if OpenSEO Belongs in My Cerebro Stack
The other half of the OpenSEO story above — what it costs to actually run it, and what I’m measuring before I decide it stays.
My favourite MCPs: macos-automator
Some software you just can’t reach. Well, macos-automator gets it.
Upgrades
What actually shipped between 28 August and 3 September, and whether you can use it. Compiled from each vendor's own changelog and release notes.
OpenAI
GPT-6 Astra
OpenAI's most capable model, and the first it rates Critical for cybersecurity capability under its own Preparedness Framework — meaning, in OpenAI's words, that it can find previously unknown security flaws and develop new ways to exploit them across well-protected systems without a person guiding each step. 1,050,000-token context, knowledge cutoff 30 April 2026. (announced, not generally available)
Who can use it: Enterprises in OpenAI's Trusted Access Program today. API, Plus, Pro, Business and Enterprise are stated as coming 'in the coming days', with no date. Free is not mentioned. The safety page calls Astra 'the most capable model we have ever broadly deployed' while the release notes say it is 'not yet generally available'.
Daybreak for Frontline Defenders
A $1bn commitment to subsidise access to OpenAI's Daybreak cyber models, plus training, technical support and partnerships, for defenders of essential services — water, electricity, local government, banking. Includes a pilot with MS-ISAC and 35+ products via a Daybreak Defense Network. (announced, not generally available)
Who can use it: Not open enrolment and not priced. Restricted to frontline defenders of essential services; the announcement's 'Work with us' section is the only stated route in. No eligibility threshold given.
Multiple Google accounts
Attach more than one Gmail, Calendar and Contacts account, so work and personal appear in one conversation.
Who can use it: Plus, Pro, Business, Enterprise. Not Free.
Site tools in the desktop browser
ChatGPT can use tools a website publishes via WebMCP, without you connecting anything.
Who can use it: ChatGPT Work and Codex, desktop app's built-in browser only — not the Chrome extension.
Browser extension reaches Edge, Brave, Opera, Vivaldi
Mention open tabs as context or hand ChatGPT a browser task from those browsers.
Who can use it: Desktop app users. Side chat works in Edge, Brave and Vivaldi but not Opera.
Google
Gemini 3.8 Flash
New model for coding and multi-step agents, at 3.7 Flash's speed and price.
Who can use it: Consumers on AI Pro and Ultra — not the free tier. Developers at $0.75/$3.75 per million, rising to $1.50/$7.50 on 1 Jan 2027.
Agentic video understanding
The model picks which parts of a video to watch instead of sampling every frame, cutting token use by up to 88%.
Who can use it: Developers now, in AI Studio and the Gemini API. Standard token pricing.
Gmail Live, Docs Live, Keep Live
Talk to Gmail to find things, to Docs to draft by voice, to Keep to turn spoken notes into structured ones.
Who can use it: Gmail and Keep Live on AI Plus, Pro and Ultra. Docs Live on Pro and Ultra only. Workspace business not yet.
Gemini 3.8 Flash Cyber
Security model for finding and patching vulnerabilities. (announced, not generally available)
Who can use it: Not generally available. Fairwind programme only — governments, national cyber authorities, critical infrastructure and core platforms.
Anthropic
Claude Fable 5.1
New top-end Claude model for coding, knowledge work and long-running agents.
Who can use it: API and the cloud resellers. On claude.ai: not on Free; Pro via usage credits, Max, Team and Enterprise included.
Cache reads cut 75%
Re-reading context you have already sent costs $0.25 per million tokens. Anthropic estimates ~25% off typical bills and ~45% off heavy agent workloads.
Who can use it: Anyone billed by token. Input and output prices unchanged.
GitHub
Copilot code review can approve pull requests
Copilot judges whether a change is ready and, if an admin allows it, formally approves it toward required-approvals.
Who can use it: Public preview on Pro, Pro+, Max, Business, Enterprise. Approval is off by default.
NVIDIA
PAIR, a personal AI router
Finds other PCs on your home network and spreads AI work across them instead of queuing on one graphics card.
Who can use it: Free and open-source beta. Windows, macOS, Linux. Needs RTX 20-series or newer, an RTX PRO GPU, a DGX Spark, or an Apple M4 or newer. Works with Ollama and LM Studio.
Agreement to acquire Hugging Face
NVIDIA has agreed to acquire Hugging Face for $12,930,300,000. Not closed. NVIDIA states the platform stays open and that its compute will not be required to build or deploy there.
Who can use it: No change to Hugging Face today.
Adobe
Adobe for Slack
Ask Slackbot to make images, video, designs and PDFs with 70+ Adobe tools, using files already in the channel.
Who can use it: Slack Business+ and Enterprise+ only.
Cursor
Self-hosted machines for Cloud Agents
Cursor's cloud agents run commands and edit files on machines you control, so code and secrets stay in your network.
Who can use it: My Machines on individual accounts. Team Pools require Enterprise.
DeepSeek
DeepSeek-V4-Flash-Vision-Exp
DeepSeek's first experimental model that reads images as well as text, published as downloadable weights.
Who can use it: Open weights, ungated, MIT licence (~168GB). Paid API endpoint also available.
Microsoft
AI text predictions now off by default
Typing suggestions no longer appear in Word and Outlook unless you switch them on.
Who can use it: Word web, Windows, Android, iOS and classic Outlook. Applies to free and paid alike.
Mistral
OCR 4.1 generally available
Document OCR with paragraph-level bounding boxes, block labels and per-block confidence scores.
Who can use it: Mistral API. A Premier commercial model — pay-per-use; self-deployment needs a commercial licence.
Mozilla
Firefox AI Controls
One switch turns off every AI feature in Firefox, present and future, with individual toggles beneath it.
Who can use it: Free, all Firefox users.
Perplexity
Hybrid compute on Mac
Splits a task between a cloud model and one running on your Mac, with an on-device gate that masks sensitive file content before anything leaves.
Who can use it: Pro, Max and Enterprise. Apple silicon, macOS 15+, 24GB unified memory minimum.
Vercel
Per-user spending budgets on AI Gateway
Caps what each team member can spend on models per day, week or month, so one runaway agent cannot drain the budget.
Who can use it: Needs CLI 59.6.2+. Keys made before this release are attributed to the team, not the person — rotate them or budgets will not apply.
Going away, or changing under you
Anthropic — tool_choice values removed
On Fable and Mythos 5.1 the values any and tool now return a 400. Existing code using them breaks.
Affects: Anyone calling those models.
OpenAI — Sora API shuts down
The Videos API and every sora-2 model are removed on 24 September. OpenAI lists no replacement.
Affects: Anyone building on it. The consumer app already closed on 26 April.
GitHub — Six Copilot models deprecated
Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Sonnet 4.5 and 4.6, and Raptor Mini removed. Sonnet 4.6 survives only for individual annual subscribers.
Affects: All Copilot users.
Cloudflare — D1 free-tier limits now enforced
The free database stops serving queries for the rest of the day once you pass the daily allowance, instead of warning. Data is untouched; queries resume at midnight UTC.
Affects: Everyone on the Workers Free plan.
Raycast — Bring Your Own Model moves behind Pro
Custom providers, local models via Ollama and OpenRouter keys now require a Pro subscription.
Affects: Existing users on the free tier lose it.
Suno — Download limits, applied retroactively
Free accounts get 7 lifetime downloads, Pro 20/month, Premier 60/month — and the limits apply to songs made before 3 September.
Affects: All tiers. New terms of service the same day.
Pinecone — Egress metering introduced
Monthly read allowances: 1GB Starter, 10GB Builder, 100GB Standard and Enterprise. On Starter and Builder, reads returning record data are blocked once spent.
Affects: All plans. The one-time $250 / 1TB import credit is withdrawn for new subscriptions.
Read the full upgrade log
Jim
