# get merge: Wednesday, September 30, 2026

27 stories merged from 4 newsletters.

AI & design news, merged daily. One briefing that merges and deduplicates stories from leading AI and design newsletters.

Web: https://getmerge.ai/2026-09-30

## TLDR

- OpenAI's DevDay led with value: GPT-6.1 Sol matches Astra on coding at roughly a fifth of the cost, cached input drops to $0.10 per million tokens, and Dots brings always-on agents with their own cloud computers to ChatGPT, part of 20+ launches that also include a hosted computer-use API, Plugin Extensions, a Marketplace and a $500/month Ultrafast tier.
- Trump rejected new AI regulation and signed the White House Accord on Super Intelligence with six lab leaders, a “morally binding” code of internal controls, external audits and board oversight that carries no legal enforcement.
- OpenAI's annualized revenue neared $70B as it weighs a $30B raise at a $1.4T valuation, Meta shipped an agent for small businesses, and a semi-humanoid robot spent an hour doing hotel laundry on camera, uncut.

## Top Stories

- **OpenAI launches Dots, always-on agents with their own cloud computers**. DevDay landed with 20+ launches, and the headline was Dots, OpenAI's answer to Meta's Muse and Grok-bot: always-on agents with their own cloud computer and browser, 4,000+ plugins and multiple ongoing projects. A Dot works across apps like Slack, Messages and Teams, runs on GPT-6 Astra, and takes action once a goal is defined. It remembers the goal, notices changes, keeps working and returns when needed. Users set Custom Rules for what it can do on its own, what needs permission and what it can't do, and sensitive actions like changing a password or deleting data need explicit approval. It starts with one Dot per user, with teams of Dots planned, and it's available starting today for Pro and Business Premium users on web, mobile and desktop. Business, Enterprise and Edu Dots are not used to train OpenAI's models by default; personal plans need an extra step to turn that off. The Neuron's framing: ChatGPT becomes the interface while the machinery fades into the background; its DevDay watch party closed on the line “the browser is being consumed by the agent window.” The Deep View flags the standing risk: agents that go rogue tend to slip past the very safeguards. Sources: The Neuron, The Deep View. [1](https://openai.com/index/introducing-dots/) [2](https://www.thedeepview.com/articles/openai-unveils-dots-at-devday-to-redefine-agents) [3](https://www.youtube.com/live/AMZj-lvIfo4?si=JyQDw6IYLFA_1MSe&t=7254)
- **GPT-6.1 Sol brings near-Astra coding at about a fifth of the price**. GPT-6.1 Sol closes most of the gap to flagship Astra on coding, document analysis and multi-step tasks without a price increase: $2 per million input tokens, $0.10 per million cached, $10 per million output, near-Astra performance at roughly a fifth of the cost. The per-task numbers: DeepSWE coding at $0.65 against Astra's $3.92, OSWorld computer use within 2 points of Astra at one-seventh the cost, Terminal-Bench science work at $5.47 against $23.80. Cached input is half of what GPT-6 Sol charged, a 95% discount off standard input pricing, and The Neuron's Corey called the caching math “what makes agents doable.” It's live in the API as gpt-6.1-sol, plus ChatGPT Work and Codex. Sources: AlphaSignal, The Neuron, The Deep View. [4](https://openai.com/index/introducing-gpt-6-1-sol/) [5](https://www.youtube.com/live/AMZj-lvIfo4?si=JyQDw6IYLFA_1MSe&t=5605)
- **Everything else DevDay shipped, from computer use to a $500 Ultrafast tier**. The rest of the stack: hosted computer use in the Agents API, so agents get a browser desktop to click and type through; Plugin Extensions, which give outside apps panels, forms, viewers and actions inside ChatGPT, with Adobe and Figma in the launch sidebar; Sign in with ChatGPT, so participating apps draw on a Plus or Pro allowance without API keys; and the OpenAI Marketplace, where eligible US enterprises apply existing OpenAI spend toward approved third-party software, in beta with 32 partners including Adobe, Figma, Sierra and Harvey. Also shipped: ChatGPT Space, a shared workspace for you, teammates and AI; Codex in the cloud, moving work between desktop, mobile and web; Private Intelligence, with zero data retention and private safety processing for automated reviews, plus a Private Inference preview; and the Decisions API, which takes a question plus a list of answers and picks one in 150ms on GPT-6 Luna, against 1.6s for a regular Luna call, limited to selected API customers for now. The trade at the top: a new $500/month tier with Ultrafast (up to 8x Codex speed on GPT-6 Astra, 6x in the API), while Pro's included Work and Codex usage drops from 20x to 10x the Plus allowance and GPT-6 Pro messages go from 200 to 100 per week. The Deep View reads the day as aimed at Anthropic's power users more than Meta's consumers. Sources: AlphaSignal, The Neuron, The Deep View. [6](https://openai.com/index/devday-2026-recap/) [7](https://developers.openai.com/api/docs/guides/agents-api/tools/computer-use) [8](https://developers.openai.com/plugins/build/extensions) [9](https://help.openai.com/en/articles/20001542-using-your-chatgpt-plan-in-other-apps-and-sites) [10](https://openai.com/business/marketplace/) [11](https://developers.openai.com/api/docs/guides/private-safety-processing)
- **The White House puts AI safety on the honor system**. President Trump met AI leaders on Tuesday, rejected the idea that the industry needs new regulation, and signed the White House Accord on Super Intelligence with six of its biggest names: Anthropic's Dario Amodei, Google's Sundar Pichai, Meta's Mark Zuckerberg, OpenAI's Greg Brockman, Elon Musk and Nvidia's Jensen Huang. Trump called the code “morally binding.” It commits each company to four layers: internal controls monitoring capability and alignment during training and deployment, including cybersecurity, biosecurity and chemical risks; an internal team with authority over safeguards and monitoring; external auditors; and a board committee overseeing the whole stack. The signatories will keep meeting to set standards, and the accord allows that it “may make sense” to codify the steps into law over time. Earlier this month Trump had called AI safety alarmism a “hoax” on stage with Huang. The Deep View notes the only enforcement is the honor system, and self-accountability gets harder as Anthropic and OpenAI move toward public listings. Sources: The Deep View. [12](https://www.politico.com/news/2026/09/29/ai-execs-morally-binding-constitution-01097855) [13](https://www.bloomberg.com/news/articles/2026-09-29/trump-set-to-host-nvidia-anthropic-ceos-to-discuss-ai-risks)
- **OpenAI nears a $70B run rate and weighs a $30B round at a $1.4T valuation**. OpenAI's annualized revenue run rate neared $70B after more than 70% growth since the start of Q3, per Axios, and it's reportedly in talks to raise $30B at a $1.4 trillion valuation, per TechCrunch. Sources: The Neuron. [14](https://www.axios.com/2026/09/29/scoop-openais-annual-recurring-revenue-nears-70b) [15](https://techcrunch.com/2026/09/29/openai-reportedly-in-talks-to-raise-30b-round-at-1-4t-valuation/)
- **Meta ships Muse for Small Business**. Meta launched Muse for Small Business, an agent that handles background work across connected business apps and requires approval before it publishes, sends or spends. It extends Meta's enterprise push while the consumer Muse is still answering for its privacy record. Sources: The Neuron, The Deep View. [16](https://www.cnbc.com/2026/09/29/meta-launches-muse-small-business-zuckerberg-pushes-enterprise-ai.html) [17](https://about.fb.com/news/2026/09/introducing-muse-small-business/)
- **A semi-humanoid robot spent an hour doing hotel laundry, uncut**. Dyna Robotics put Dyna-2.1 on Taku, a semi-humanoid robot, and filmed about an hour of hotel laundry without cutting the awkward parts: dozens of small decisions and manipulations instead of one pre-scripted fold for a 20-second highlight reel. Real work is a chain of boring edge cases: grab this towel, move that pile, recover when something lands weird, figure out what comes next, repeat for an hour. The Neuron's favorite benchmark is whether it can survive a fitted sheet. Sources: The Neuron. [18](https://www.dyna.co/dyna-2.1) [19](https://x.com/DynaRobotics/status/2104970523726033387)

## Deeper Cuts

- **Design reading: AI is a boring technology, and how to keep enjoying programming**. An essay at chrbutler.com, “AI is a boring technology,” argues AI is less interesting than what it helps you do, and means that as a compliment. The tech gets good when it fades into infrastructure, like electricity. A companion read on the Haskell forum, “How to keep enjoying programming in a world of LLMs,” starts from the fear of losing your job to someone with little skill, no interest in quality and a huge Claude account. Sources: Sidebar. [20](https://www.chrbutler.com/ai-is-a-boring-technology) [21](https://discourse.haskell.org/t/how-to-keep-enjoying-programming-in-a-world-of-llms/14705)
- **Design finds: sidenotes without JavaScript, WebGPU for agents, and a hacker simulator**. Sidenotes in the margin with CSS anchor positioning and no JavaScript, falling back below the text on narrow viewports and older browsers. vgpu.sh is a WebGPU library documented for agents: fullscreen shaders, compute pipelines, raw WebGPU interop, and read-only source files compiled by the docs app itself. And an online hacker simulator, for when you want to pretend to hack a network. Sources: Sidebar. [22](https://vincent.bernat.ch/en/blog/2026-css-sidenotes) [23](https://vgpu.sh/examples) [24](https://hacker-simulator.com/)
- **Sonnet 5.5 matches Opus 5.5 on agentic tasks while using 7x the tokens**. One caveat on Sonnet 5.5's price story: Artificial Analysis has it matching Opus 5.5 on agentic tasks while using about 7x more tokens, so the savings on long agent runs are thinner than the sticker price suggests. Sources: AlphaSignal. [25](https://artificialanalysis.ai/models/releases/claude-sonnet-5-5)
- **Meta's training trick doubles Qwen3-8B's math accuracy and cuts output length**. Meta's self-evolving training trick doubles Qwen3-8B's math accuracy while cutting output length, and the write-up is on arXiv. Sources: AlphaSignal. [26](https://arxiv.org/abs/2609.30652)
- **Make the model restate your goal before it starts**. Lauren Tan's most-used prompt, worth pasting before a complicated research, coding, planning or writing task: “restate in your own words what you think my goals are and what the problem I'm trying to solve is.” A model can execute the wrong interpretation perfectly, and the restatement surfaces the mismatch before you burn time, tokens or 14 tool calls. Sources: The Neuron. [27](https://x.com/poteto/status/2104744961904394699)
- **Anthropic says GLM-5 can autonomously build cyber exploits**. Anthropic published research on GLM-5 autonomously building cyber exploits, part of its report on the spread of advanced cyber capabilities. Sources: The Deep View. [28](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities)
- **Business notes: McDonald's ML pricing, Gemini Gems to Skills, Apple's AppleCare cuts**. McDonald's used machine learning across millions of restaurant tickets to recommend local menu prices from willingness-to-pay and competitor data, with franchisees keeping final control. Google says Gemini Gems will automatically migrate to Skills for personal accounts in November, replacing standalone custom assistants with reusable instructions Gemini can auto-apply or stack in any chat. Apple reportedly planned to cut 5,000 AppleCare roles for agents. And OpenAI faces a lawsuit over the Hugging Face breach. Sources: The Neuron, The Deep View. [29](https://www.reuters.com/business/inside-mcdonalds-push-have-ai-price-your-big-mac-2026-09-29/) [30](https://support.google.com/gemini/answer/18560919?hl=en) [31](https://www.macrumors.com/2026/09/29/apple-reportedly-planned-5000-applecare-layoffs/) [32](https://www.axios.com/2026/09/29/openai-sued-hugging-face-breach)
- **Science notes: fusion math with Astra, Isomorphic's drug search, and models passing traits**. OpenAI's GPT-6 Astra helped researchers find new families of plasma equilibria in fusion math. Isomorphic Labs' drug-design agent searched huge chemical spaces for Pareto-best tradeoffs, with no disease target or clinical candidate named yet. And researchers found AI models can transfer behavioral traits through apparently unrelated training data, which raises questions about hidden model-to-model influence. Sources: The Neuron, The Deep View. [33](https://theneuron.ai/news/gpt-6-astra-fusion-math-breakthrough-grads-conjecture/) [34](https://www.isomorphiclabs.com/articles/building-a-new-path-to-make-medicines-with-ai) [35](https://theneuron.ai/explainer-articles/ai-models-may-be-able-to-teach-each-other-skills-without-talking-about-them/)

## Tools and Launches

- **PageIndex**. Tree search over documents instead of vector chunks: no vector database, no embeddings, no chunking. It builds a tree structure from a document and reasons through it, the way a human flips to the right chapter, and it scores 98.7% on FinanceBench, ahead of every vector RAG on the leaderboard. Open source, self-hostable, pip install, with MCP and API support for production. Sources: AlphaSignal. [36](https://github.com/VectifyAI/PageIndex)
- **RedAmon**. An open-source framework that runs full cyberattacks and auto-patches your code. Sources: AlphaSignal. [37](https://github.com/samugit83/redamon)
- **CYBER-FROST 3.8 GGUF**. Blackfrost AI's 180B MoE model with 6B active parameters, in GGUF format for local inference. Sources: AlphaSignal. [38](https://huggingface.co/freakyskittle/CYBER-FROST-3.8-GGUF)
- **airi**. An open-source AI companion that runs locally, plays Minecraft, and supports 30+ LLM models. Sources: AlphaSignal. [39](https://github.com/moeru-ai/airi)
- **OpenClaw Enterprise**. Multi-tenancy, permissions, auditing and swappable model and sandbox layers for persistent agents in sensitive environments. Sources: The Neuron. [40](https://openclaw.ai/blog/openclaw-enterprise)
- **Liquid d1**. Decision models that return yes/no, a choice, or a scored decision with probabilities instead of spending tokens generating prose. Sources: The Neuron. [41](https://docs.liquid.ai/lfm/models/decision-models)
- **InstaCloud**. Serverless compute, Postgres, branching environments and deploys that coding agents operate end to end through CLI and skills. Sources: The Neuron. [42](https://www.instacloud.com/)
- **OpenResearch**. Turns coding agents into experiment runners with isolated git worktrees and an immutable experiment tree. Sources: The Neuron. [43](https://github.com/alphaXiv/OpenResearch)
- **Perplexity Automations**. Ongoing work and actions in response to triggers or schedules, working with memory, skills and connected apps. Sources: The Deep View. [44](https://x.com/perplexity_ai/status/2104976036274552963)
- **DiffusionGemma**. Google's open-source model can now be turned into a “Jev-like” decision model. Sources: The Deep View. [45](https://x.com/googlegemma/status/2104990261181075498)
- **Google Diffusion Controller**. A lightweight steering damper that precisely steers image generation for better prompt alignment. Sources: The Deep View. [46](https://x.com/GoogleResearch/status/2105005817103823011)
- **America.gov**. The US government's new chatbot answers typed or spoken federal-service questions from official sources and can strip personal details from uploaded forms before processing them. Sources: The Neuron, The Deep View. [47](https://america.gov/)

## References

1. [OpenAI: Introducing Dots](https://openai.com/index/introducing-dots/)
2. [The Deep View: OpenAI unveils Dots at DevDay to redefine agents](https://www.thedeepview.com/articles/openai-unveils-dots-at-devday-to-redefine-agents)
3. [YouTube: The Neuron's DevDay watch party: the agent window](https://www.youtube.com/live/AMZj-lvIfo4?si=JyQDw6IYLFA_1MSe&t=7254)
4. [OpenAI: Introducing GPT-6.1 Sol](https://openai.com/index/introducing-gpt-6-1-sol/)
5. [YouTube: The Neuron's DevDay watch party: caching makes agents doable](https://www.youtube.com/live/AMZj-lvIfo4?si=JyQDw6IYLFA_1MSe&t=5605)
6. [OpenAI: DevDay 2026 recap](https://openai.com/index/devday-2026-recap/)
7. [OpenAI: Computer use in the Agents API](https://developers.openai.com/api/docs/guides/agents-api/tools/computer-use)
8. [OpenAI: Plugin Extensions](https://developers.openai.com/plugins/build/extensions)
9. [OpenAI Help: Sign in with ChatGPT](https://help.openai.com/en/articles/20001542-using-your-chatgpt-plan-in-other-apps-and-sites)
10. [OpenAI: Marketplace](https://openai.com/business/marketplace/)
11. [OpenAI: Private Safety Processing](https://developers.openai.com/api/docs/guides/private-safety-processing)
12. [Politico: AI execs sign a “morally binding” accord](https://www.politico.com/news/2026/09/29/ai-execs-morally-binding-constitution-01097855)
13. [Bloomberg: Trump hosts Nvidia and Anthropic CEOs on AI risks](https://www.bloomberg.com/news/articles/2026-09-29/trump-set-to-host-nvidia-anthropic-ceos-to-discuss-ai-risks)
14. [Axios: OpenAI's annual recurring revenue nears $70B](https://www.axios.com/2026/09/29/scoop-openais-annual-recurring-revenue-nears-70b)
15. [TechCrunch: OpenAI reportedly raising $30B at a $1.4T valuation](https://techcrunch.com/2026/09/29/openai-reportedly-in-talks-to-raise-30b-round-at-1-4t-valuation/)
16. [CNBC: Meta launches Muse for Small Business](https://www.cnbc.com/2026/09/29/meta-launches-muse-small-business-zuckerberg-pushes-enterprise-ai.html)
17. [Meta: Introducing Muse for Small Business](https://about.fb.com/news/2026/09/introducing-muse-small-business/)
18. [Dyna Robotics: Dyna-2.1](https://www.dyna.co/dyna-2.1)
19. [X: Dyna Robotics: an uncut hour of hotel laundry](https://x.com/DynaRobotics/status/2104970523726033387)
20. [chrbutler.com: AI is a boring technology](https://www.chrbutler.com/ai-is-a-boring-technology)
21. [Haskell Discourse: How to keep enjoying programming in a world of LLMs](https://discourse.haskell.org/t/how-to-keep-enjoying-programming-in-a-world-of-llms/14705)
22. [Vincent Bernat: Sidenotes with CSS anchor positioning](https://vincent.bernat.ch/en/blog/2026-css-sidenotes)
23. [vgpu.sh: A WebGPU library designed for agents](https://vgpu.sh/examples)
24. [Hacker Simulator: Online hacker simulator](https://hacker-simulator.com/)
25. [Artificial Analysis: Claude Sonnet 5.5](https://artificialanalysis.ai/models/releases/claude-sonnet-5-5)
26. [arXiv: Meta's training trick doubles Qwen3-8B math accuracy](https://arxiv.org/abs/2609.30652)
27. [X: Lauren Tan on restating your goal first](https://x.com/poteto/status/2104744961904394699)
28. [Anthropic: GLM-5.3 and the spread of advanced cyber capabilities](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities)
29. [Reuters: Inside McDonald's push to have AI price your Big Mac](https://www.reuters.com/business/inside-mcdonalds-push-have-ai-price-your-big-mac-2026-09-29/)
30. [Google: Gemini Gems moving to Skills](https://support.google.com/gemini/answer/18560919?hl=en)
31. [MacRumors: Apple reportedly planned 5,000 AppleCare layoffs](https://www.macrumors.com/2026/09/29/apple-reportedly-planned-5000-applecare-layoffs/)
32. [Axios: OpenAI sued over Hugging Face breach](https://www.axios.com/2026/09/29/openai-sued-hugging-face-breach)
33. [The Neuron: GPT-6 Astra fusion math breakthrough](https://theneuron.ai/news/gpt-6-astra-fusion-math-breakthrough-grads-conjecture/)
34. [Isomorphic Labs: Building a new path to make medicines with AI](https://www.isomorphiclabs.com/articles/building-a-new-path-to-make-medicines-with-ai)
35. [The Neuron: AI models may teach each other skills without talking about them](https://theneuron.ai/explainer-articles/ai-models-may-be-able-to-teach-each-other-skills-without-talking-about-them/)
36. [GitHub: VectifyAI/PageIndex](https://github.com/VectifyAI/PageIndex)
37. [GitHub: samugit83/redamon](https://github.com/samugit83/redamon)
38. [Hugging Face: CYBER-FROST 3.8 GGUF](https://huggingface.co/freakyskittle/CYBER-FROST-3.8-GGUF)
39. [GitHub: moeru-ai/airi](https://github.com/moeru-ai/airi)
40. [OpenClaw: OpenClaw Enterprise](https://openclaw.ai/blog/openclaw-enterprise)
41. [Liquid AI: LFM decision models](https://docs.liquid.ai/lfm/models/decision-models)
42. [InstaCloud](https://www.instacloud.com/)
43. [GitHub: alphaXiv/OpenResearch](https://github.com/alphaXiv/OpenResearch)
44. [X: Perplexity Automations](https://x.com/perplexity_ai/status/2104976036274552963)
45. [X: Google Gemma on DiffusionGemma](https://x.com/googlegemma/status/2104990261181075498)
46. [X: Google Research on the Diffusion Controller](https://x.com/GoogleResearch/status/2105005817103823011)
47. [America.gov: US government chatbot](https://america.gov/)

Sources today: [AlphaSignal](https://alphasignal.ai/), [The Neuron](https://www.theneurondaily.com/), [Sidebar](https://sidebar.io/), [The Deep View](https://www.thedeepview.co/).