OpenAI Ships GPT-6.1 Sol, Dots Agent, and Ultrafast Tier as White House Superintelligence Accord Is Signed

Key Takeaways

1. OpenAI released GPT-6.1 Sol at one-fifth of Astra's price with a 95% cache-read discount, plus the always-on Dots agent and an Ultrafast tier up to 8x faster in Codex.
2. The White House hosted a superintelligence meeting where Trump renamed AI to "superintelligence" and frontier lab leaders signed the White House Superintelligence Accord and Joint Frontier Responsibility Commitment.
3. SpaceX's Starship flew its 14th mission, reaching orbit for the first time and deploying 26 Starlink V3 satellites.
4. NVIDIA open-sourced Physis-Lang and Kumo Tabular, and joined Red Hat and OpenAI on OpenClaw Enterprise.
5. Anthropic launched a new qualitative research effort through Anthropic Interviewer, open to Claude and Claude Code users from September 29 to October 6.
6. Hugging Face released d1, its first decision model, and TaskSmith, a harness for generating RL environments from code.
7. Runway joined the OpenAI Marketplace as a launch partner and introduced Agent Tagging.

1. OpenAI's Model and Product Wave

1. OpenAI released GPT-6.1 Sol, which it says delivers intelligence close to Astra at one-fifth the price, with a 95% cache-read discount; Sam Altman called it "the absolute workhorse model." Sol is coming soon with Ultrafast, while Astra is available today. 1 2
2. OpenAI also launched Dots, an "always-on" agent powered by GPT-6 Astra that is positioned to handle a wide range of tasks. Altman described Dots as a new 24/7 way to use AI that returns time and attention to higher-level work. 1 2
3. OpenAI introduced the Ultrafast speed tier, claiming up to 8x faster token generation in Codex (300 tokens per second) and up to 6x faster in the API. Altman said ultrafast is so fast he never wants to go back. 1 2
4. Codex Security Cloud received a major upgrade, now including access to network-capable models through Daybreak Blue by default. It can scan entire GitHub repositories, continuously review new commits, investigate and deduplicate findings, and prepare fixes for review, running even when the computer is off, delivered as Codex desktop and web plugins. 1
5. Altman said the team completely rewrote ChatGPT Web on top of the desktop app codebase over the past five weeks, which he said was only possible with Astra's support and is why Dots, Space, Codex Cloud and other releases landed on web and desktop the same day. He called it the start of the "giant project era." 1
6. Ethan Mollick, after briefly using Dots, placed it alongside Muse and Grokbot in a fast-expanding "Clawlike" category, saying capable models with access to personal data are useful as assistants and second opinions, but he could not give a detailed review due to short use. He also criticized OpenAI for repeatedly launching similar third-party ecosystems and abandoning them: Plugins in 2023, GPTs and the GPT Store in 2023-2024, Apps in 2025, and a differently defined Plugins in 2026. He noted that Dot, Spaces, Pages, local/cloud ChatGPT Work, and cloud and desktop scheduled tasks now coexist, making the overall picture confusing and overlapping. 1 2 3

2. White House Superintelligence Meeting and Industry Commitments

1. The White House held a superintelligence meeting attended by Trump and tech leaders. Trump announced that "artificial intelligence" would officially be renamed "superintelligence," said the US leads in AI and will maintain that lead, said the industry could exceed the Industrial Revolution in scale, and emphasized self-regulation alongside government oversight. 1 2
2. Frontier lab leaders signed the White House Superintelligence Accord and the Joint Frontier Responsibility Commitment, accepting responsibility for safe product development and introducing internal controls and external audits. Musk said this is better than waiting for an international agreement that may never come. 1 2
3. Musk and Jensen Huang described the accord as establishing a joint oversight committee and specialized committees, with companies "grading each other's homework" rather than self-assessing. Huang said the accord includes strict internal controls, upgraded internal audits, third-party external audits, and sharing of industry best practices. 1
4. Demis Hassabis welcomed progress on the White House superintelligence accord and said he looks forward to follow-up. A quoted post said the author met with the US President, Vice President, House Speaker, and government and tech leaders that day and signed the accord; it also said Google has invested hundreds of billions of dollars in related areas over the past two years and will increase that, covering the entire technology stack, and that Google emphasizes testing, evaluation, and red-teaming before releasing models or products. 1
5. Musk said AI is most likely to bring an "age of abundance," meaning universal high income rather than just basic income, with robots and superintelligence able to give everyone in the world better medical care than anyone present. He said employment roles will evolve and that demand for human labor is currently enormous. 1
6. Grok powers the US government's new AI website, according to US Chief Design Officer Joe Gebbia, who said the site is driven by Grok and Gemini so the public can get answers from official government sources. Musk congratulated Gebbia. Separately, Grok 4.7 xHigh ranked first on the Artificial Analysis Cyber Index, ahead of Fable 5.1 Max, Opus 5.5, Astra 6, and GPT-6, positioned as a frontier model for enterprise cyber defense. 1 2 3

3. Infrastructure, Open Source, and Developer Tools

1. SpaceX's Starship flew its 14th mission and reached orbit for the first time: Ship 41 arrived in orbit on September 28, and although one Raptor engine shut down early, SpaceX used another burn to complete orbital insertion. It deployed 26 Starlink V3 satellites at about 275 km altitude, the first operational payload in orbit. SpaceX successfully deployed Starlink V3 satellites and established contact for the first time; at scale, a single Starship could carry 60 V3 satellites, with network capacity roughly equivalent to 20 Falcon 9 launches. 1 2
2. Starlink Mobile launched in Ecuador with CNT_EC, connecting cellular dead zones including the Amazon rainforest and the Galápagos Islands. A new Roadster demonstration was postponed by two weeks to October 15 due to high winds, as the event can only be held outdoors. 1 2
3. NVIDIA released Physis-Lang, an open-source self-evolving framework that adds physical reasoning to video descriptions, explaining why and how scenes unfold. These descriptions are used to fine-tune world models and to add reasoning to prompts during video generation. Adding physical reasoning to prompts alone, without retraining, improved NVIDIA Cosmos 3's PhyGenBench score by 5.62 points. NVIDIA also introduced Kumo Tabular, a new family of foundation models for tabular data that it says establishes a new Pareto frontier in the accuracy-versus-inference-time tradeoff, released with open weights, open-source software, and a permissive commercial license. 1 2
4. The OpenClaw Foundation partnered with Red Hat, NVIDIA, and OpenAI to open-source OpenClaw Enterprise, an enterprise control plane for persistent agents that can run on an organization's own infrastructure and is permanently free for organizations. NVIDIA said organizations can use NVIDIA OpenShell as an open-source option to govern agents in OpenClaw Enterprise. NVIDIA also proposed agentic development on Jetson as a new way to build edge AI, and demonstrated with NousResearch/Teknium using NVIDIA NeMo Relay to collect traces for Hermes Agent: running two example scenarios, tracking agent calls and retries, and viewing full traces in Arize Phoenix. 1 2 3
5. Hugging Face released d1, its first decision model, which it says surpassed Jev on the Hugging Face Decision Index for the first time. Advantages include multilingual evaluation, resistance to prompt injection, and longer input handling, aimed at fast structured decisions in software environments; it is available via Liquid API and coming soon to OpenRouter. Hugging Face also released TaskSmith, a specialized harness for generating RL environments from code, described not as a general coding agent with long prompts but as a dedicated orchestrator whose goal is to turn PRs into the best possible RL environments. 1 2
6. Runway joined the OpenAI Marketplace as a launch partner, giving users access to its image, video, audio, and editing models, with some OpenAI committed credits usable for Runway purchases. Runway introduced Agent Tagging, letting users tag the Runway Agent anywhere in their material to request edits, options, and fixes. Runway also shared a workflow for building an Ad Studio with Runway Dev and coding agents: turning one product photo into a full social campaign, generating campaign images, adding and translating on-image copy, and exporting product and UGC videos for websites, TikTok, and Reels. 1 2 3

4. Research, Evaluation, and Practitioner Notes

1. Anthropic launched a new research effort through Anthropic Interviewer to understand users' experiences with AI, what role they want AI to play in their personal lives and the world, and their expectations of AI companies. Last December, 81,000 people participated in what Anthropic calls its largest qualitative study, sharing hopes and concerns about AI; this time participants can choose to make answers public so people outside Anthropic can access them. Anthropic said feedback will influence research by The Anthropic Institute and its decisions, and if many people believe Anthropic and other companies should change their practices, those views will be recorded and citable by anyone. The study runs September 29 to October 6 and is open to Claude and Claude Code Free, Pro, and Max users. 1
2. Hamel Husain offered four approaches for AI evaluation when traces contain sensitive data: negotiate scoped access, work with authorized experts, de-identify, and use synthetic data. He recommended efficient sampling of production traces for review: explore broadly first, then use signals to locate likely failures, and keep some random traces in each batch to discover new issues. He attended OpenAI DevDay in San Francisco, live-blogged the keynote, and reacted strongly to "Sign in with ChatGPT," noting it lets users log into third-party apps and use their already-paid tokens. 1 2 3 4
3. Deedy shared a video-generation workflow after more than 10 hours optimizing how to use Opus 5.5, recommending Claude Code rather than the app itself and using the OpenRouter API with one key to call image, video, and audio models. The workflow includes using Gemini 3.8 TTS for emotional voiceover as a skill; Manim, Hyperframes, and Motion Canvas for HTML web-video motion graphics; GPT 2.5 Image Sunburst for keyframes; and Veo 3.1 / Seedance 2.5 for video, with Seedance better at motion shots. The author stressed preparing reference images for consistency and generating an animatic before the full video, and recommended adding a scriptwriting planning skill, assembling with OpenTimelineIO, generating background music with ElevenLabs, generating word-level timeline subtitles with an ASR model, and adding a critic skill to verify audio and video quality via screenshots and transcripts. Prompts should specify desired content, aspect ratio, duration, and style preferences, with most actual video processing done by ffmpeg; the author also advised explicitly avoiding "Claudisms" such as short sentences and many numbers, instead narrating "like a university professor," and allowing yt-dlp to search and inject real Creative Commons video clips. 1
4. Yann LeCun shared a JEPA-based world model, AD-E2E-JEPA, for end-to-end autonomous driving, whose core question is whether a world model can drive without training a driving policy; the work was done with Kevin Ghstz and Professor Anna Choromanska. He also shared a historical analogy: in 1501, Pope Alexander VI criticized the safety of Gutenberg's printing press, saying printing was greatly beneficial for spreading useful and tested books but could also cause serious harm, making comprehensive control necessary. 1 2
5. Ethan Mollick argued that, regardless of Anthropic's motives for publishing its research, open-weight models will soon create the same kinds of safety threats already demonstrated by closed models, with the difference being the absence of guardrails, and that "we are very close," so planning ahead is advisable. He also argued that one feature of the AI business is that labs with frontier models can ship half-finished products that still perform surprisingly well because the AI can figure things out and improvise, effectively embedding a forward-deployed engineer and customer support in the product. 1 2
6. Aravind Srinivas compared companies to cars that become autonomous with the right automation, saying Perplexity Computer can do this today but that determining the right automation still requires significant human skill and initiative. A quoted post said Perplexity Computer's Automations are for persistent work, triggered by events or schedules and working with memory, skills, and connected apps such as Slack, Gmail, Outlook, and Linear. Another quoted view said elite long-short analysts and growth investors from the author's Goldman Sachs days do not use AI for Q&A but for scheduled or event-triggered tasks: activating 15-20 automations at market close and 3-5 when a CEO files a new document or posts on X, covering and analyzing large domains with under 5 minutes of latency. Srinivas also said Computer can create interactive 3D diagrams to help visualize things like furniture assembly, and that the only remaining step is projecting this while wearing a pair of glasses. 1 2 3 4