Deep Dive
Dots: Autonomous AI agents that learn your context
Sam Altman opens by introducing Dots, the keynote's flagship product. These are always-on AI agents built on Astra that handle tasks 24/7 without requiring users to ask for each step. The demo features Alfred, a personal dot that juggles work and life: updating website designs, preparing board decks showing revenue up 51% year-over-year, mapping code migrations, and even booking wedding cake tastings. What differentiates Dots from past virtual assistants is scope and agency. Instead of booking reservations, Dots tackle ambitious multi-day projects like migrating a legacy API across an entire codebase by tracing dependencies, writing code, running tests, and submitting pull requests for review. They come with their own cloud computer, browser, and code execution environment. Users stay in control via customizable boundaries and instructions, and Dots integrate with 4,000+ ChatGPT apps already connected. The rollout strategy prioritizes Pro, Business Premium, and Enterprise tiers, with no usage charges eating into plan limits. For enterprises, specialist Dots are in preview—company-wide virtual teammates for accounting, legal, and marketing that learn from feedback and scale work across teams.
ChatGPT Spaces: Collaborative workspaces for teams and agents
Complementing individual Dots, OpenAI introduced Spaces—live collaborative pages where teams and agents work together. Think of it as a Notion-meets-Figma-meets-Slack hybrid built inside ChatGPT. Users can write plans, generate data, create interactive charts, and tag Dots or teammates directly in comments to hand off work. Holly's demo showed a fictional music app launch where her Dot Doy automatically caught overnight updates, flagged design revisions, and built features based on voice commands. The key innovation is dynamic content: Dots can keep charts and FAQs updated hourly, and the interface stays agent-readable. Presentations are coming soon directly inside Spaces, designed so agents can easily edit them. Holly also emphasized voice as her personal favorite—she uses phone booth time to brainstorm with Doy, which already knows her context and can execute immediately. Slack integration lets her forward threads to her Dot for full delegation, freeing her to focus on strategy. The demo revealed unexpected behavior: OpenAI employees began treating Dots like delegates and inviting them to group chats from the start, turning them into team extensions rather than individual assistants.
New models: GPT-6.1 Soul and Ultra Fast for speed and cost
On the model front, Sam announced two new releases responding to the most common developer requests: cheaper and faster. GPT-6.1 Soul hits near-Astra capability at one-fifth the cost, with 95% discounts on cached input tokens—a game-changer for agents reusing context across long tasks. Ultra Fast delivers eight times the speed of standard models at 300 tokens per second, six times the cost, but worthwhile for real-time applications. Raman's demo showed Ultra Fast building and rendering 3D worlds from sketches in seconds, retraining game graphics from rough to polished in a few turns, and even playing browser games autonomously by learning controls on the fly. Sam also previewed the Decisions API, which constrains outputs to predefined options, achieving sub-millisecond inference for classification, routing, and agentic decisions. This unlocks robot control and real-time browser automation that felt impossible before. Pricing tiering expanded: Pro 500 offers 25x the usage of Plus with Ultra Fast access, and Pro 200 reopens with GPT-6.1 Soul as the daily driver, giving users best-in-class intelligence at every price point.
AI as research intern: models solving decade-old math problems
Tel, a research lead on post-training, revealed how Astra is accelerating OpenAI's own research. Models now help optimize the computer use harness itself—running loops to find latency and performance wins that shipped as 2x improvements. On safety, models helped improve monitoring and refusal training, pushing Astra toward industry-leading performance on computer use stress tests. The breakthrough metric: in January, models succeeded on short research tasks but failed almost all day-long tasks. By July, with Ultra Fast, models completed over one-third of day-long research tasks with zero human intervention. This enables scientific work at scale: Astra has solved over 100 math problems open for decades, helped create antibiotics for drug-resistant infections, and accelerated renewable energy development. Tel emphasized this virtuous cycle—better models help train even better aligned models, which loop back into more capable products for developers.
Developer tools: Codex in cloud, Agents API, and private intelligence
Sam detailed the platform layer for builders. The Codex harness, powering Codex work and Dots, is now open source and fully cloud-based, letting tasks persist across devices without keeping laptops open. Raman's demo spanned use cases: building 3D worlds with voice in Ultra Fast, using the refreshed CLI with bidirectional audio, running app migrations and code refactoring in the cloud, and—most striking—connecting to a HuggingFace programmable robot (MicroDuck named Lavender) that Codex wired up, giving it vision, voice, image generation, and game-playing abilities in hours. The Agents API entered public beta with multi-agent controls and browser/computer use capabilities. For security, the new CodeQL Security Cloud scans for vulnerabilities in cloud environments with verified fixes. OpenAI also previewed Private Intelligence with private safety processing and private inference, letting frontier models work on sensitive data without storing it on OpenAI servers. Performance gains across the board: 100x growth in API responses in one year while maintaining 99%+ reliability, 45% faster time-to-first-token, 30% faster tool calls. AWS partnership lets customers build OpenAI agents on Bedrock alongside existing applications.
Developer monetization: marketplace, sign-in, and plugin extensions
To help developers build businesses, OpenAI launched three mechanisms. First, Sign in with ChatGPT lets users bring their existing subscription tokens to third-party apps—no token cost friction for onboarding. Second, Plugin Extensions make apps feel native to ChatGPT and Codex: Meetings app shows calendar and notes in-chat, Figma lets designers iterate in-conversation, Adobe brings Photoshop tools directly into ChatGPT. These plugins can now appear conversationally—if a user asks a question, ChatGPT proactively suggests relevant plugins without manual search. Third, the OpenAI Marketplace (30+ launch partners: Notion, Code Rabbit, Burcel) lets enterprise customers spend existing commitments on third-party products, creating a direct path to B2B revenue. ChatGPT Sites now support plugins and authenticated users, enabling personalized multi-user experiences. The submission process simplified: developers can track reviews, request human help, and update tools without restarting approval. This ecosystem aims to help thousands of builders reach 1.2 billion ChatGPT users.