OpenAI DevDay 2026 Keynote - CourtneyR.dev's 2nd Brain

OpenAI DevDay 2026 Keynote ![rw-book-cover](https://d34adp677peecb.cloudfront.net/static/images/article1.be68295a7e40.png) 🔗 **Source:** [OpenAI DevDay 2026 Keynote](https://www.youtube.com/watch?v=Fls_onRviPM) > [!note]+ Document Notes > tags: [openai-devday, agent-platforms, model-pricing, decisions-api, codex-cloud, plugin-distribution] > author: Sam Altman, Romain Huet, Tejal Patwardhan, Holly Li > source: OpenAI (YouTube) > reference: https://www.youtube.com/watch?v=Fls_onRviPM > date created: 2026-09-29 16:30 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: OpenAI DevDay 2026 Keynote (FULL) >  ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--openai-devday-2026-keynote-full) ## 📝 Summary - **What happened.** OpenAI's official channel published the 53:34 DevDay 2026 keynote on 2026-09-29 (180,086 views at capture). The video description names Sam Altman, Romain Huet, Tejal Patwardhan and Holly Li and promises "20+ launches." The video has no chapter list. Every product, price, number and availability date below is a vendor claim from the company that sells the products. Treat each as unverified until the docs, pricing page or a third-party test confirms it. - **Caption garbles.** The auto-captions render the flagship model as "Astra" (and once "Astro"), the mid-tier model as "GPT 6.1 Soul" (the description says "GPT-6.1 Sol"), the fast tier as "ultra fast" / "ultraast" (the description says "Astra Ultrafast"), Codex as "codeex" / "codecs," the OpenAI product name as "chatbt" / "Chat GBT," and the presenters as "Tel," "Tachel," "Raman," "Ramon" and "Roma." This analysis uses the description's forms where they exist and preserves the caption spelling inside quote blocks. - **The claim.** OpenAI ships a product layer ("dots," always-on agents with their own cloud computer, plus "ChatGPT Spaces" for shared pages) and a developer layer (a cheaper GPT-6.1 Sol model, an 8x-speed "Ultrafast" tier, a preview "decisions API," Codex in the cloud, computer use in the agents API, private-inference controls, plugin extensions, a marketplace, and "Sign in with ChatGPT"). - **Why it matters.** For developers, the durable pieces are the pricing tiers (a claimed fifth of the flagship price, a claimed 95% cache-input discount), the closed-set decisions API, and the distribution changes (plugin extensions, in-conversation discovery, a marketplace payable from enterprise commitments). Each changes what you build against and where users find it. - **Key stakeholders.** OpenAI (vendor of everything shown), ChatGPT Pro, Business and Enterprise subscribers, API developers, AWS (Bedrock managed agents partnership), Microsoft (Agent 365 integration of specialist dots), Hugging Face (loaned the robot), Figma and Adobe (plugin extension examples), and the 16 launch partners for Sign in with ChatGPT. - **Risks and tradeoffs.** Four live demos stalled on stage (voice at 00:11:54, a Codex thread error at 00:15:33, voice again at 00:33:00 and 00:35:22). No benchmark, eval name or methodology appears for the "most aligned," "industry-leading" and "over a third of day-long research tasks" claims. Distribution features (in-chat discovery, marketplace) also route customers and commitments through OpenAI. - **Bottom line.** Useful as an inventory of what OpenAI says it shipped on 2026-09-29. Not evidence that any claim holds. Verify pricing, limits, region availability and the decisions API's semantics on the docs before building. ## 📖 Main content **Structure.** The keynote runs in four blocks. Consumer and prosumer products (00:01:05 to 00:18:32), models (00:18:32 to 00:27:14), developer tools (00:27:14 to 00:43:28), and distribution (00:43:28 to 00:47:53). A wrap-up, a video of builders, logistics for the rest of the day, a joke "bank reset," and a closing statement on AI and work (00:47:53 to 00:52:54) finish it. **Dots.** The pitch is an always-on agent that "knows what matters to you" and handles delegated work without a prompt each time. Dots have their own cloud computer and browser, follow your access, and work with "more than 4,000 apps" through the plugins you already connected. The opening ran as a scripted video (00:02:37 to 00:05:25), then Holly Li ran a live demo with a fictional music app called Blossom. Text and phone-call interfaces are "coming soon." You start with one dot and may later run a team. Dots run on Astra and ship with app and computer-use boundaries and custom instructions. **Spaces.** A shared page where people and dots edit together. You can tag a dot in a comment, give a page standing instructions (for example, update it daily from a Slack channel), and insert charts, sheets and live prototypes with a slash menu. A presentation format "easily read and modified by agents" is announced as coming soon. **Availability and pricing, as stated.** Dots and Spaces are available today to Pro, Business and Enterprise users, and dot conversations "don't eat into your usage." Specialist dots (company-configured virtual teammates for accounting, marketing, legal) are an Enterprise preview. Pro 500 is a new tier with 25 times the usage of Plus and Ultrafast access. Pro 200 reopens. **Models.** GPT-6.1 Sol is pitched at "very near Astro level intelligence at a fifth of the price," with a 95% discount on cached input. Fast is 2x speed at 2x price. Ultrafast is 8x speed at 6x price, quoted at 300 tokens per second, on Astra now and on Sol later. The decisions API (preview) gives a "Luna" model a predefined option set and returns in a fraction of a second. OpenAI says it tested this with computer use "not sped up." **Research claims.** Tejal Patwardhan says an "AI research intern" goal from a live stream a year earlier has now been hit. Models optimized the computer-use harness in a loop (a claimed 2x-plus latency win, shipped), improved refusal behavior for computer use, and completed "over a third" of day-long research tasks with no intervention by July, up from mostly failing in January. Named outside results (over 100 open math problems, antibiotics work) come with no citations in the transcript. **Developer tools.** The Codex harness is open source and now powers Codex and dots. Codex runs fully in the cloud with phone-to-desktop handoff. Codex Security in the cloud (public beta) scans for vulnerabilities and prepares verified fixes, with deduplication and scheduled scans. Computer use comes to the agents API. Bedrock managed agents powered by OpenAI arrive on AWS. "OpenAI private intelligence" (preview) combines zero data retention with private safety processing and private inference. Responses API claims: 100x growth in a year, above 99% reliability, 45% lower time to first token, more than 30% faster tool calls. **Demos by Romain Huet.** Ultrafast edits a 3D world, an Ultrafast CLI builds and modifies a raffle app (three attendees, then six), Astra plays a game through browser use, "appshots" give Codex app context, a cloud task ("rewrite the entire back end in Rust") starts from the phone, and a Hugging Face programmable robot combines vision, image generation (GPT image 2.5) and voice (GPT live one) with the decisions API named as the route for near-real-time robot actions. **Distribution.** About 1.2 billion people use ChatGPT weekly (vendor claim). Sign in with ChatGPT lets users spend their plan's tokens in your app, with 16 launch partners. Plugin extensions let you ship editors, dashboards and workspaces inside ChatGPT and Codex, with Figma and Adobe as examples among "dozens" of launch partners. Plugins also work in ChatGPT sites, discovery moves into the conversation, and the submission process gains review tracking and human escalation. The OpenAI marketplace has 30-plus launch partners and lets enterprises pay with existing OpenAI commitments, with open-source models available through a partner the caption renders "B 10." **Contested or unverifiable in this source.** - "Most aligned model in the market" and "industry-leading performance on computer use stress tests" name no eval. - "Nothing like this on the market today" (Ultrafast) is a comparison with no named competitor. - The Ultrafast side-by-side (a rocket in Dev Day colors) is a single visual prompt. - Price multiples are relative to "standard" with no dollar figures spoken. - The 1.2 billion weekly users, 4,000 apps, 100x Responses API growth and 99% reliability figures carry no source. ## 🗒️ Notes #### ⏱️ [00:00:22](https://www.youtube.com/watch?v=Fls_onRviPM&t=22) Welcome and shipped-from-Twitter requests - Opening frames DevDay as OpenAI's "best dev day yet" and credits developer requests, mostly on Twitter. - Three named requests already shipped: Codex on Linux, multiple folders in one Codex project, Codex on a phone. - Evidence is a spoken list. No usage numbers for any of the three. #### ⏱️ [00:01:05](https://www.youtube.com/watch?v=Fls_onRviPM&t=65) The dots pitch - Positioned against "virtual assistants" that book reservations, which the presenter calls trivial by comparison. - Framing device: an AI "from the movies" that knows what matters to you and works 24/7. - Commercial interest: the pitch sells a subscription-bundled product, so the "always on" framing is also a retention argument. #### ⏱️ [00:02:37](https://www.youtube.com/watch?v=Fls_onRviPM&t=157) Meet dots (scripted video) - Scripted sequence: a dot named "Alfred" updates a website, a board deck (revenue "up 51% year-over-year" in the fictional scenario), an app migration PR, and personal errands. - Every result is pre-scripted, so it demonstrates the interface, not accuracy. - The user approves publishing and rollout steps, which shows the intended human checkpoint. #### ⏱️ [00:05:25](https://www.youtube.com/watch?v=Fls_onRviPM&t=325) Delegation, cloud computer and controls - Example responsibilities: monitor bug reports and draft PRs, run a budget planning cycle, find what slows an app and show the speedup. - Legacy API migration example: trace dependencies, write code, run tests, open a PR for review. - Each dot gets its own cloud computer and browser and follows the user's access. Text and voice-call channels are "coming soon." > [!quote] [00:07:35] > Dots are powered by Astra, our most aligned model, and they have built-in protections and safeguards. #### ⏱️ [00:07:35](https://www.youtube.com/watch?v=Fls_onRviPM&t=455) ChatGPT Spaces - Problem statement: most productivity software was not built for people and agents to work together. - A Space starts as a page. Dots track dependencies, respond to comment tags, and follow page-level instructions such as a daily Slack summary. - Agent-readable presentations are announced as coming soon. #### ⏱️ [00:09:45](https://www.youtube.com/watch?v=Fls_onRviPM&t=585) Holly Li demo: dots in a launch week - Fictional app "Blossom Music," launch "tomorrow." The dot moved a calendar event, summarized test-user feedback and surfaced a late design change. - A "go build it" handoff kicks off a Codex build on the presenter's laptop. - Voice mode stalls at 00:11:54 ("Still checking"), and the presenter routes around it by asking for an FAQ in the page. #### ⏱️ [00:14:07](https://www.youtube.com/watch?v=Fls_onRviPM&t=847) Dots inside Slack and group chats - Forwarding a thread delegates it. Dots get their own identity, so teammates treat them as delegates. - Internal claim: engineers have dots fixing "dozens of bugs a day," and dots were partly "built by dots." No count, repo or acceptance rate is given. > [!quote] [00:15:33] > Our engineers have their dots fixing dozens of bugs a day in this way - A Codex thread errors during the live demo and the presenter retries. The app build then runs in the iPhone simulator through Codex. #### ⏱️ [00:17:01](https://www.youtube.com/watch?v=Fls_onRviPM&t=1021) Availability and specialist dots - Dots and Spaces ship today to Pro, Business and Enterprise. Dot conversations do not count against usage limits, per the presenter. > [!quote] [00:17:01] > Both dots and space will be available to all chatbt pro, business premium, and enterprise users today. - "Business premium" is likely a caption garble of the plan tiers. Check the pricing page for the exact plan names. - Specialist dots (Enterprise preview) cover accounting, marketing and legal. A Microsoft Agent 365 integration is announced. #### ⏱️ [00:18:32](https://www.youtube.com/watch?v=Fls_onRviPM&t=1112) Platform framing: models, tools, distribution - Three parts: models (intelligence, speed, cost), tools (the same ones OpenAI uses internally), distribution (discovery and customers). - Astra launched "a few weeks ago"; the ask from developers was cheaper and faster. #### ⏱️ [00:19:58](https://www.youtube.com/watch?v=Fls_onRviPM&t=1198) GPT-6.1 Sol and Ultrafast - Sol: "very near" Astra intelligence at a fifth of the price, with a 95% cached-input discount, aimed at agent loops that reuse context. > [!quote] [00:19:58] > the cache input costs are a 95% discount on standard input. - Fast is 2x speed at 2x price. Ultrafast is 8x speed at 6x price, 300 tokens per second, in the API, ChatGPT and Codex. > [!quote] [00:20:46] > ultraast takes that to eight times the speed at six times the price of standard. - Demo: the same "create and launch a rocket" prompt on Ultrafast and regular side by side. One prompt, no eval. #### ⏱️ [00:21:31](https://www.youtube.com/watch?v=Fls_onRviPM&t=1291) Subscription tiers - Pro 500 is new, with Ultrafast in ChatGPT and Codex and use across partner apps through Sign in with ChatGPT. Pro 200 reopens with Sol as the daily driver. > [!quote] [00:21:31] > Pro 500 will have 25 times the usage of plus #### ⏱️ [00:22:14](https://www.youtube.com/watch?v=Fls_onRviPM&t=1334) Decisions API (preview) - A "Luna" model picks from a predefined option set: route a request, classify an image, choose the next agent step. - Claims retained image understanding, broad language support and safety protections. > [!quote] [00:22:14] > the decisions API lets the model respond in a fraction of a second is so fast. - No latency number, price or option-set limit is given. The demo shows computer-use steps "not sped up." #### ⏱️ [00:22:56](https://www.youtube.com/watch?v=Fls_onRviPM&t=1376) Research intern and Tejal Patwardhan segment - Claim: the "AI research intern" prediction from a year earlier is met. Tejal Patwardhan (caption "Tel") leads post-training research. - Computer-use harness loop found a claimed 2x-plus latency improvement that shipped, and safety monitoring changes. > [!quote] [00:25:07] > this has resulted in already a more than 2x latency win that we've shipped to all of you. - Internal token use "skyrocketed" after summer. Day-long tasks: mostly failed in January, over a third done with no intervention by July. > [!quote] [00:26:32] > our models have taken on over a third of dayong research task with no intervention required. - Named outcomes (100-plus open math problems, antibiotics, ancient languages, energy, robotics) are asserted without citations. #### ⏱️ [00:27:14](https://www.youtube.com/watch?v=Fls_onRviPM&t=1634) Codex harness, cloud Codex and Codex Security - The Codex harness is open source and now runs Codex and dots. > [!quote] [00:27:57] > Codex is now fully in the cloud. - Codex Security in the cloud (public beta): continuous scans, deduplication, scheduled scans, a new interface, verified fixes for review. > [!quote] [00:28:40] > It continuously finds vulnerabilities and prepares verified fixes for you to review. #### ⏱️ [00:29:24](https://www.youtube.com/watch?v=Fls_onRviPM&t=1764) Agents API, AWS, private intelligence, Responses API - Computer use comes to the agents API. Demo concept: an agent that opens a browser and tests a site flow. - AWS: Bedrock managed agents powered by OpenAI, with frontier models, Codex and ChatGPT work available to AWS customers. - Private intelligence (preview): zero data retention with private safety processing, plus private inference. Storage and inference privacy terms need the docs. - Responses API: 100x growth in a year, 99%-plus reliability, 45% lower time to first token, 30%-plus faster tool calls. > [!quote] [00:30:50] > We've seen 100x growth in the responses API in the last year. #### ⏱️ [00:31:33](https://www.youtube.com/watch?v=Fls_onRviPM&t=1893) Romain Huet demo: Ultrafast worlds and the CLI - A pre-recorded Codex exchange welcomes the audience, then Romain Huet (captions "Raman," "Ramon," "Roma") recreates the venue as a 3D world from photos. - Voice control fails at 00:33:00 and again at 00:35:22. The presenter switches to typing and the terminal. - Refreshed Codex CLI with bidirectional audio "powered by GPT live." Raffle app built, then changed from three to six winners in seconds. #### ⏱️ [00:36:49](https://www.youtube.com/watch?v=Fls_onRviPM&t=2209) Astra Adventures and appshots - Sketch to textured 3D game across modalities, then Astra plays the game through a browser-use connection. - "Appshots" attach an app's screen as context. Codex then drives the iOS simulator to audit screen sizes and capture screenshots. #### ⏱️ [00:40:26](https://www.youtube.com/watch?v=Fls_onRviPM&t=2426) Codex cloud tasks and the robot demo - Cloud tasks run with the same tools as the laptop app, including plugins and computer use. Demo prompt: rewrite a back end in Rust. Result is not shown. - Hugging Face "MicroDuck" robot named Lavender combines Astra vision, GPT image 2.5 and GPT live one. Decisions API is cited as the route to near-real-time robot actions from vision input. #### ⏱️ [00:43:28](https://www.youtube.com/watch?v=Fls_onRviPM&t=2608) Distribution: Sign in with ChatGPT and plugin extensions - 1.2 billion weekly ChatGPT users (vendor figure). > [!quote] [00:44:13] > they can sign into your app and use the tokens included in their chatbt plan. - 16 launch partners for Sign in with ChatGPT. Plugin extensions turn plugins into full in-ChatGPT and in-Codex apps. Examples: a meetings app, Figma design review, an Adobe Photoshop feature. - ChatGPT sites now support connected plugins and per-user credentials. The presenter says millions of sites were built within months of launch. #### ⏱️ [00:47:03](https://www.youtube.com/watch?v=Fls_onRviPM&t=2823) Plugin discovery, submission and marketplace - Plugins get suggested in the flow of a conversation. Submission gets review tracking, fix guidance, a human-review request and tool updates without resubmitting. > [!quote] [00:47:03] > You'll be able to track review, see what needs fixing, request a human, and update your plugins tools without starting the whole submission over. - Marketplace: 30-plus launch partners (captions: "Code Rabbit, notion, burcell"). Enterprises can pay with existing OpenAI commitments, and developers with a commitment can spend it there. > [!quote] [00:47:03] > our enterprise customers can now use part of their existing OpenAI commitment to buy your products. #### ⏱️ [00:47:53](https://www.youtube.com/watch?v=Fls_onRviPM&t=2873) Recap and builder video - Recap list: dots, Spaces, Sol, Ultrafast, decisions API, agents API, Codex in the cloud, business tools. - A montage video credits builders working on immune-system research and education access. #### ⏱️ [00:50:04](https://www.youtube.com/watch?v=Fls_onRviPM&t=3004) Rest of the day and the "reset" - Announced sessions: open-source investment including "our new openclaw enterprise harness," a live Q&A, a Codex game studio session, limited-edition swag for attendees. - A staged "bank reset" of usage limits, with a countdown at 00:52:10. The transcript gives no numbers for the reset and the caption spelling "banked" may be "rate limit." #### ⏱️ [00:52:10](https://www.youtube.com/watch?v=Fls_onRviPM&t=3130) Closing statement - The closing rejects the "industrial revolution" framing in favor of a "renaissance." > [!quote] [00:52:10] > There are some parts of life that we cannot and should not automate. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. The description contains no URLs. - [OpenAI DevDay 2026 Keynote (FULL), OpenAI (YouTube)](https://www.youtube.com/watch?v=Fls_onRviPM): the source video, 53:34, uploaded 2026-09-29. The description credits Sam Altman, Romain Huet, Tejal Patwardhan and Holly Li and names dots, ChatGPT Spaces, GPT-6.1 Sol and Astra Ultrafast. - OpenAI API documentation and pricing pages (platform.openai.com, openai.com/pricing): the primary source for every price multiple, model name, limit and availability date in this analysis. Not linked in the video. - OpenAI Codex and the open-source Codex harness: named at 00:27:14. The repository location is not stated in the transcript. - AWS Bedrock managed agents powered by OpenAI: named at 00:29:24. Confirm terms on AWS documentation. - Microsoft Agent 365: named at 00:17:48 as the integration point for specialist dots. - Hugging Face: loaned the "MicroDuck" programmable robot shown at 00:41:10. - Figma and Adobe: plugin extension launch partners named at 00:45:39. - The partner the caption renders "B 10" (00:47:03) supplies open-source models to the marketplace. The likely correct name is Baseten. That is an inference, so verify on OpenAI's marketplace page. - The earlier vault analysis of the Type Safe "Jev" video ([[Jev Is Not an LLM (And That's Why It Wins)]], `PNbbMi2h9VA`) covers a closed-set decisions model from a different vendor, useful as a comparison for the decisions API. ## 🗣️ Paraphrase OpenAI used its developer conference to launch two kinds of things. The first is a product for people: dots, agents that run on their own cloud computer, take delegated work, and appear in Slack, in shared pages called Spaces, and eventually by text and phone call. The second is a stack for developers: a cheaper model close to the flagship, an 8x-speed tier at 6x the price, a fast option-picking API in preview, Codex that runs entirely in the cloud, privacy controls, and three distribution channels (Sign in with ChatGPT, plugin extensions and a marketplace). The company backs the launches with internal claims. Its engineers have dots fixing dozens of bugs a day, its research models finished more than a third of day-long research tasks unaided by July, and the API grew 100x in a year while holding above 99% reliability. None of the claims comes with a benchmark, a repo or a method. The live demos also stumbled four times, and the presenters said so and moved on. The practical signal for builders is in the distribution block. Plugins can now become full apps inside ChatGPT and Codex, get suggested mid-conversation, and be bought through enterprise commitments, which puts OpenAI between developers and customers. ## 📰 Newsletter **OpenAI announced 20-plus things on 2026-09-29. Five matter if you build.** First, GPT-6.1 Sol: OpenAI says it delivers near-flagship intelligence at a fifth of the price, with cached input discounted 95%. If that holds, agent loops that reuse context get cheaper. Second, Ultrafast: 8x speed at 6x the price, 300 tokens per second, on the flagship now. Third, the decisions API (preview): a model picks from options you define and answers "in a fraction of a second." That's closed-set classification as an API, and it echoes the "classify before you generate" argument from the Type Safe Jev video we covered earlier. Fourth, Codex fully in the cloud, plus Codex Security in public beta. Fifth, plugin extensions: your plugin can become a full app inside ChatGPT and Codex. The caveat is that every number is OpenAI's own. No benchmark names appear for "most aligned" or "industry-leading," the Ultrafast comparison is one rocket prompt, and four live demos stalled on stage. For WordPress and Pantheon audiences, the plugin-extension and marketplace pieces matter most because they change where users discover tools. Read the docs before you plan around any date. Links: [the keynote](https://www.youtube.com/watch?v=Fls_onRviPM). ## ✅ Action items #### DevRel - Developer advocacy: write a "what changed on 2026-09-29" explainer that lists each launch, its status (shipped, preview, coming soon) and the doc URL, with pricing as dollars from the pricing page, not the keynote's multiples. - Education: run an office-hours session on "closed-set decisions versus generation," using the decisions API preview next to the Type Safe Jev analysis as the two case studies. - Developer experience: test Sign in with ChatGPT and plugin extensions as an integration path for one Pantheon-hosted demo app and record the setup time. - Developer marketing: do not repeat "a fifth of the price," "8x speed" or "100x growth" in Pantheon material without the pricing-page or docs citation. Label each as an OpenAI claim. - Community management: add a standing "vendor claim, verify" note to link-sharing channels for launch-day threads. - Internal/external: ask OpenAI contacts which of the "coming soon" items (text and call for dots, Ultrafast on Sol, agent-readable presentations) have dates. #### OSPO - Policy and governance: decide whether the open-source Codex harness is acceptable in contribution tooling and write the rule before anyone wires it into a repo. - License compliance: read the harness license and the private-intelligence terms before any internal or customer data touches either. - Security management: cloud agents with their own computer, browser and inherited user access are a new credential surface. Add "dot follows your access" to the threat model. - Software inventory: record any Codex cloud or Codex Security use as a hosted dependency with an availability expectation, not a library. - Developer support: publish guidance on scoping agent access (which connectors, which repos) before a team enables dots. - Community engagement: track which plugin-extension partners appear after the 2026-09-29 launch, since distribution through one vendor is a lock-in risk. - Strategic alignment: set a KPI for review time and defects per agent-authored PR, so "dozens of bugs a day" becomes measurable locally. #### Community - Maintainers: pilot Codex Security scans on one repo and compare findings with the existing scanner for 30 days, counting false positives. - Contributors: document a rule for disclosing agent-authored PRs, since dots open PRs under their own identity. - Sponsors: if a sponsor funds an AI-triage pilot, publish cost and accuracy before and after. - Ecosystem: run the same issue set through the decisions API and a standard model so results are comparable across projects. #### WordPress - Core: no action. This is vendor tooling, not core code. - Core Performance: none directly. The Responses API latency claims are network-bound and do not affect PHP request budgets. - Core AI: compare the decisions API's option-set interface with the Abilities API and any AI client guidance. Note that the preview's semantics (out-of-set handling, confidence output) are undocumented in the keynote. - Core Program Team: flag that "specialist dots" and "Spaces" are product names, not established categories, in cross-team notes. - Mobile: none. - Playground: a browser demo that runs one closed-set classification through the decisions API and one through a general model would show the tradeoff once the API is public. - Community: none. - Training: draft a lesson on delegating to agents with scoped access, using the dots controls (app and computer boundaries, custom instructions) as the example. - Marketing: none. - Support: evaluate the decisions API for ticket routing once it exits preview, measured against current human routing. - Design: none. - Documentation: add an "agents that follow your access" security note to any AI docs page, and track plugin extension docs for WordPress-relevant patterns. - Photos: none from this source. - TV: none. - Polyglots: the decisions API claims "broad language support" without a language list. Ask for one before testing localized routing. - Accessibility: none from this source, but plugin extensions render WordPress-adjacent UI inside ChatGPT and should be checked against WCAG 2.2 AA when they ship. - Meta: none. - CLI: the refreshed Codex CLI with bidirectional audio is worth a hands-on test for WP-CLI workflows, after the voice reliability shown on stage improves. - Hosting: ChatGPT sites (millions built, per OpenAI) and cloud Codex overlap with hosting use cases. Track connected plugins and per-user credentials on those sites as a competitive and integration signal. - Tide: Codex Security's deduplication and scheduled scans map to Tide's finding triage. Compare outputs on a known-labeled set. - Openverse: none from this source. - Themes: none. - Plugins: read the plugin extension and marketplace submission changes (review tracking, human review request, tool updates without resubmission) and compare them with WordPress.org's plugin review flow for ideas. #### The WPOCC - Programs: none directly. If The WP Open Community Collective funds AI-assisted maintainer tooling, inference cost per resolved decision is a reasonable line item. - Projects: none from this source. - Funding experiments: if a microgrant covers agent-assisted triage, require published before and after cost and accuracy numbers. ### FAIR - Technical independence: dots, cloud Codex and the decisions API are hosted, closed services. FAIR-adjacent tooling that depends on them inherits a single-vendor availability dependency. Document that before adoption. - Policies and covenants: require agent decisions in any package-provenance workflow to log the decision, the input and the confidence, so the outcome is auditable. - Productized governance artifacts: add a "hosted agent dependency" row to operator scorecards next to hosted models, covering data retention (ZDR or not) and where inference runs. ## 🧪 Hypotheses 1. GPT-6.1 Sol reaches at least 90% of Astra's score on a coding task suite at 25% or less of Astra's cost per completed task, which is looser than the "fifth of the price" and "very near" claims. 2. The cached-input discount (95% on standard input) cuts total cost per completed agent task by more than 40% on workloads that reuse a 20,000-token or longer context across five or more steps. 3. Ultrafast delivers 8x wall-clock speed but costs 6x per task, so cost per unit time is lower while cost per completed task is higher for short tasks. 4. The decisions API returns a correct option from a fixed set with accuracy within 2 points of a general model on a 500-item labeled routing set, at more than 5x lower latency. 5. Dots opening agent-authored PRs at "dozens of bugs a day" produce a merge rate below 60% on external repos with no dedicated triage, because the internal figure comes from a curated feedback channel. 6. Plugin discovery in the flow of a conversation sends more first-time installs than library search within 90 days of launch. ## 📊 Methodology - **Design for H1 and H2.** Pick 100 tasks from an open coding benchmark that OpenAI has not named in the keynote. Run each on Astra and Sol with the same harness and cached context. Record dollars from the invoice, tokens by type (uncached input, cached input, output) and pass rate. Report cost per passing task with confidence intervals. - **Design for H3.** Time 50 tasks of three lengths (under 1 minute, 5 minutes, 30 minutes) on Standard, Fast and Ultrafast. Compare wall-clock time and cost per completed task. Measure time to first token separately from tokens per second. - **Design for H4.** Assemble a 500-item labeled set from support tickets or issue labels. Run the decisions API and a general model with a schema. Report accuracy, macro-F1, calibration if confidence is exposed, and p50 and p95 latency from three network locations. - **Design for H5.** Enable a dot on three test repos with seeded bugs. Measure PRs opened, PRs merged after human review and reviewer minutes per PR. Repeat with no curated channel. - **Design for H6.** Ship a small plugin extension, log the install source (in-conversation suggestion, library search, direct link) and compare counts over 90 days. - **Controls.** Fix prompts, seeds and hardware. Pre-register pass thresholds before running. State the model version string from the API response, since the captions garble names. ## ❓ Questions - What eval or benchmark supports "most aligned model in the market" and "industry-leading" computer-use safety, and where is it published? - What are the dollar prices for Sol, Fast and Ultrafast per million input, cached-input and output tokens? The keynote gives only multiples. - Does Ultrafast bill 6x on input, output or both, and does it apply to cached input? - What does the decisions API do when the correct answer is not in the option set, and does it return a confidence? - What data does a dot retain, and does the "conversations don't eat into your usage" policy apply to API and Codex use inside a dot? - How does "ZDR with private safety processing" differ from standard zero data retention, and what does private inference change on the wire? - Which plan tiers exactly received dots and Spaces "today," given the "business premium" caption garble? - What is the revenue split, review timeline and commitment-drawdown rule for the OpenAI marketplace? ## 🔮 Future research - **Agent access scope and blast radius.** Dots "follow your access" with their own browser and computer. Independent research on what happens when a connected app's content injects instructions into an always-on agent is thin, and a 24/7 agent raises the exposure window. - **Delegated PR quality at scale.** "Dozens of bugs a day" needs an external replication with merge rates, reversions and reviewer cost. - **Closed-set decision APIs as a category.** Two vendors now sell option-picking models (OpenAI's decisions API and the Type Safe Jev model). A shared benchmark for calibration, out-of-set handling and latency would make claims comparable. - **Distribution concentration.** Discovery inside the conversation, sign-in with the user's subscription and marketplace payment through commitments concentrate demand at one vendor. Studies of developer margin and switching cost would inform platform choices. - **Self-improving research systems.** The keynote reports models optimizing their own computer-use harness and finishing a third of day-long tasks. Outside verification of task selection and "no intervention" definitions is missing. ## 💡 Implications - **Practice.** Cache reuse (a claimed 95% input discount) makes long-lived context a design choice: keep stable prefixes stable. Split agent work into decisions (small, fast) and generation (large, slower) where the decisions API or a similar closed-set model fits. - **Policy.** Agents that inherit user access, act 24/7 and open PRs under their own identity need written scope, approval and disclosure rules. The keynote mentions boundaries and custom instructions but shows no audit log. - **Security.** Cloud agents with a browser and computer widen the attack surface. Codex Security (public beta) and private inference are OpenAI's answers, both vendor claims until tested. Add prompt injection through connected apps to every threat model. - **Education.** The 25x usage jump (Pro 500) and cheaper Sol lower the cost of teaching with agents, though price multiples need dollar checks. Teach delegation with scoped access as a skill. - **Funding.** Marketplace payment through enterprise commitments means a customer can buy your plugin without new procurement, which shifts sales cycles. Sponsors of open-source tooling should ask whether their commitments can fund community projects. - **Open source.** The harness is open source, but the models, dots and cloud runtime are not. That gap is where lock-in accumulates. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript body is about grade 5, and Flesch Reading Ease is about 81, from an automated count (8,924 words, 720 sentences, roughly 1.34 syllables per word). Spoken keynote English with short sentences. The comprehension load comes from product names and unstated context (Astra, Sol, Luna, Ultrafast, plugin extensions), not sentence difficulty. #### 👀 Visual learners - Present: live demos of dots, Spaces, Ultrafast versus regular side by side, a 3D world, an on-screen decision trace and a robot. - Improve: the transcript cannot confirm any pricing table, chart or latency graph on screen. Put a single pricing table on screen for Sol, Fast and Ultrafast, and read its dollar values aloud. #### 🔊 Aural learners - Present: continuous narration, audience reactions and a clear spoken structure ("three parts": models, tools, distribution). - Improve: multiplier claims ("a fifth of the price," "eight times the speed at six times the price") arrive in fast succession with no spoken dollar figure. Read the base price once so a listener can compute the rest. #### 🗣️ Verbal learners - Present: explicit definitions of what dots and Spaces are, and a recap list at 00:47:53. - Improve: "harness," "ZDR," "private safety processing," "private inference" and "appshot" go undefined. One clause each would fix it. #### 🧩 Kinesthetic learners - Present: multiple task demos (raffle app, game, audit of an app across screen sizes, cloud task). - Improve: no repo, quickstart or copy-paste snippet is given in the video. Viewers watch but cannot try the demos, and several demos stalled live. #### 🧠 Logical learners - Present: a stated structure and named metrics (time to first token down 45%, 99% reliability, 100x growth). - Improve: no denominators or baselines. 45% down from what, measured on which requests and over what period? #### 🧑‍🤝‍🧑 Social learners - Present: a live audience, a joint countdown for the usage reset, and mentions of Q&A and hands-on sessions. - Improve: no community channel or discussion thread is named for online viewers. #### 🧘‍♀️ Solitary learners - Present: a 53-minute recording that supports pausing and rewatching, with a clear recap. - Improve: no chapters and no self-check. A reflection prompt ("which of your agent steps is a decision, not generation?") would convert the launches into a planning exercise. #### 🌿 Naturalistic learners - Present: little explicit. The robot duck and the 3D venue world are playful physical-world hooks. - Improve: an 8x-speed tier at 6x price, and a 100x API growth claim, imply large compute and energy use. The keynote never mentions energy per request. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies landmarks and player controls, and the transcript is machine-readable. - Improve: the demos rely on visuals (dashboards, charts, a decision trace, a robot drawing) that narration only sometimes describes. WCAG 2.2 AA 1.1.1 (Non-text Content) and 1.2.5 (Audio Description, Prerecorded) are not met for those segments. #### Keyboard navigation - Present: the YouTube player is keyboard operable with visible focus by default. - Improve: no chapters, so keyboard users must scrub. Adding chapters would support 2.4.5 (Multiple Ways) and 2.4.6 (Headings and Labels) in spirit, though they apply to web content rather than the video itself. #### Contrast and readability - Present: unverifiable from a transcript. Product UI on stage is presumably designed for projection. - Improve: code, charts and small UI text at default YouTube resolution risk failing 1.4.3 (Contrast Minimum) and are unreadable on mobile. Zoom regions during demos. #### Text resizing and reflow - Present: the player scales and captions resize independently. - Improve: burned-in slide and UI text does not resize. 1.4.4 (Resize Text) and 1.4.10 (Reflow) fail for text baked into video frames. #### Media and captions - Present: an auto-generated transcript exists, with timestamps. - Improve: the captions garble the flagship model name ("Astro" and "Astra"), the mid-tier ("Soul" for Sol), the presenters ("Tel," "Tachel," "Raman," "Ramon"), the product ("chatbt," "Chat GBT"), the tools ("codeex," "codecs") and partner names ("B 10," "burcell"). Speaker changes are marked only sporadically. Substantively, this fails 1.2.2 (Captions, Prerecorded) for proper nouns, and 1.2.4 (Captions, Live) applied to the live stream. Human-corrected captions are the fix. #### Forms and interactions - Present: none in the video. - Improve: the products shown (voice-first dots, Spaces slash menus) need their own 2.5.8 (Target Size) and 3.3.2 (Labels or Instructions) checks. Voice-only paths such as the phone-call mode need text alternatives (the text-message mode is a start). #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate captions on proper nouns), **1.2.5** (no audio description for visual-only demo content), **1.1.1** (charts and UI shown without description), **1.4.4 / 1.4.10** (burned-in text). The captions failure is the highest-impact and cheapest to fix. ## 🌍 DEIB assessment #### Cultural breadth - Present: examples include a music app, wedding planning, kids' lessons, a 3D game, a research use case and a robot. Named outside benefits include ancient-language study and accessibility. - Improve: the demos are US and English centered (Fort Mason, San Francisco fun fact, board decks, PTA-style scheduling). "Broad language support" for the decisions API is claimed without a demo in a non-English language. #### Inclusive language - Present: no exclusionary terms in the transcript. Language is direct and second-person. - Improve: superlatives ("remarkably capable," "superpowers," "absurdly fast") pack claims with no fact behind them. Replace them with the measured figure. #### Neurodiversity support - Present: strong chunking into three named parts and a recap at 00:47:53. Repeated demos follow a predictable order of prompt, result, next step. - Improve: no chapters, and a 53-minute runtime with rapid stage transitions taxes attention. Fast, overlapping demo chatter and live failures add processing load. Publish chapters and a written summary. #### Representation - Present: four named speakers appear across the video and a montage of builders shown at 00:48:35 to 00:50:04. - Improve: the transcript does not identify the montage participants. Name them and credit their work so representation is checkable, not asserted. #### Psychological safety - Present: presenters acknowledge failures on stage ("Live demos," "We might have to type this time") without blame, which models low-stakes error handling. - Improve: framing that dots make people "feel capable" of more may pressure viewers who don't see a use case. Pair it with a plain statement of what dots don't do well yet. ## ✅ Logical validity #### Strong reasoning - **Mechanism argument, 00:19:58.** Cached input at a 95% discount makes agents that reuse context cheaper. The mechanism (reused tokens cost less) is checkable on the pricing page and holds if the discount is real. - **Product-design argument, 00:07:35.** Most productivity software predates agents, so shared workspaces need agent-readable structure. The claim is testable, and the Spaces demo (tagging a dot in a comment) shows one concrete design. - **Self-disclosure of failures, 00:11:54, 00:15:33, 00:33:00, 00:35:22.** Stalled voice, a Codex error and voice problems are acknowledged live. These are concessions against interest and raise credibility of the demos that did run. - **Tradeoff stated openly, 00:20:46.** Fast is 2x speed at 2x price and Ultrafast 8x speed at 6x price. The tradeoff is explicit and lets a listener compute a cost-per-time comparison, even though dollar figures are missing. - **Scoped claim, 00:22:14.** The decisions API works "by giving our Luna model a predefined set of options," which names its mechanism. That is a narrower, more defensible claim than "fast model." #### Fallacies - **Unsupported superlative, 00:19:14.** "the most capable and aligned model in the market" carries no eval, no comparison set and no date. *Fix:* name the benchmarks, models and dates, or state the specific measured behaviors. - **Unsupported comparison, 00:20:46.** "There is nothing like this on the market today" is a universal claim with no named alternative. Other vendors and open-source stacks also sell fast inference. *Fix:* state the tokens per second and price, then let listeners compare. - **Hasty generalization, 00:20:46.** One "rocket" prompt side by side stands for Ultrafast versus regular in general. *Fix:* report a task set with medians and spread. - **Anecdote as evidence, 00:15:33.** "Our engineers have their dots fixing dozens of bugs a day in this way" is a self-report from a curated environment. It gives no merge rate, reverted-fix rate or reviewer time. *Fix:* publish the counts and acceptance rate. - **Impressive-example bundling, 00:26:32.** Over 100 open math problems, antibiotics and ancient languages are listed as achievements of "models like Astra" with no citations, and the causal role of the model is asserted rather than shown. *Fix:* link each result and state what the model contributed versus researchers. - **Undisclosed-interest framing, throughout.** OpenAI presents its own subscription tiers, marketplace and commitments as developer benefits with no note that each also routes spend and discovery through OpenAI. *Fix:* state the commercial model next to each feature. - **Ambiguity from garbled captions, throughout.** "Astro" versus "Astra," "Soul" versus "Sol," and "business premium" versus plan names make quoted facts unusable without correction. #### Tone of voice - Enthusiastic, warm and promotional, in a keynote register. Second person, contractions and stage humor ("the reset company," "Let's do a little countdown") throughout. Superlatives cluster at the framing moments (00:01:05 to 00:02:37, 00:18:32, 00:52:10) while the model, pricing and API sections carry concrete figures. The presenters handle failed demos with light self-deprecation. The closing at 00:52:10 shifts to a values statement about automation and choice. Reading heuristic: treat the figures as claims to verify and the framing as marketing. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response pattern to accusation or criticism. This is a product keynote with no accuser and no defensive posture. For completeness across the three prongs: **Deny**: nothing is denied. The presenters acknowledge failed demos directly (00:11:54, 00:15:33, 00:33:00). **Attack**: no critic or competitor is named or disparaged. "Nothing like this on the market today" implies a comparison, but it doesn't attack anyone. **Reverse victim/offender**: no such move appears. One passage (00:52:10) rejects the "industrial revolution" framing of AI, and it does so as a values statement, not as a rebuttal to a named critic. ## 🎯 Conclusion The keynote is a launch inventory from a vendor with a commercial interest in each item. Its most decision-relevant developer claims are GPT-6.1 Sol at near-Astra intelligence for a fifth of the price, a 95% cached-input discount, Ultrafast at 8x speed for 6x the price, a decisions API for closed-set choices, Codex fully in the cloud, and three distribution channels. The reasoning holds where it states a mechanism (cache reuse, a predefined option set, an explicit speed and price tradeoff) and is thin where it states a superlative (most aligned, nothing like it on the market, dozens of bugs a day). The gap that matters most is missing dollar prices, benchmarks and availability details. The multiples are relative to a "standard" tier the video doesn't price, the safety and alignment claims name no eval, and the "coming soon" items (text and call for dots, Ultrafast on Sol, agent-readable presentations) carry no date. Four stalled live demos also show the voice-first path is not yet reliable in the conditions of a conference. **Next steps for readers:** pull the pricing page and API docs for Sol, Fast and Ultrafast and compute cost per completed task on your own workload. Test the decisions API against your labeled routing set once it is available. If you ship plugins, read the plugin extension and marketplace terms before planning distribution. Treat every "today" as a claim to confirm on the docs. Correct the caption garbles (Sol, Astra, Codex) before reusing any quote. ## 🔄 Refactor Appendix — OpenAI DevDay 2026 Keynote (FULL) > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied. The "most aligned," > research-intern and 4,000-app claims were dropped as unevidenced in this source. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Sol pitches near-flagship intelligence at a fifth of the price > [!quote] [00:19:58] > gives you very near Astro level intelligence at a fifth of the price to help you code and build. **Context & use**: - 00:19:58, the GPT-6.1 Sol launch (caption "Soul"). Vendor claim, no benchmark named and no dollar price spoken. - Keep as a test target: run Sol against Astra on your own task set before believing "very near." **Tags**: #xref #articleseed  #### [H2] Ultrafast trades 6x price for 8x speed > [!quote] [00:20:46] > ultraast takes that to eight times the speed at six times the price of standard. **Context & use**: - 00:20:46, after Fast's stated 2x speed at 2x price. Caption spells the product "ultraast." - The multiple is against an unpriced "standard." Compute cost per completed task, not per token, before adopting. **Tags**: #xref  #### [H3] The decisions API restricts a model to predefined options > [!quote] [00:22:14] > giving our Luna model a predefined set of options to choose from like routing a request, classifying image, deciding what agent should do next **Context & use**: - 00:22:14, the preview announcement. Mechanism is closed-set choice, the same shape as the Type Safe Jev model from the earlier vault analysis. - Undocumented in the keynote: out-of-set behavior, confidence output, price and limits. **Tags**: #xref #articleseed  #### [H4] Research models finished a third of day-long tasks unaided > [!quote] [00:26:32] > our models have taken on over a third of dayong research task with no intervention required. **Context & use**: - 00:26:32, from Tejal Patwardhan's segment. Task set, grading and "no intervention" definition are not given. - Teaching specimen for reading internal-progress claims: a rate without a denominator or a task list. **Tags**: #xref #share  #### [H5] Marketplace lets enterprises pay with existing commitments > [!quote] [00:47:03] > our enterprise customers can now use part of their existing OpenAI commitment to buy your products. **Context & use**: - 00:47:03, the marketplace announcement with 30-plus launch partners. - Distribution and procurement now run through the model vendor. Read the revenue split and review terms before building for it. **Tags**: #xref #articleseed  ### 🧠 Atomic Split Pad #### [P1] Read launch multiples against an unpriced baseline **Claim** (one sentence, Courtney's voice): A vendor's "a fifth of the price" or "8x speed at 6x price" tells me nothing until the baseline price, the task and the unit (per token or per completed task) travel with it. **Evidence pointer**: [[#H1]] · [[#H2]] **Scope & conditions**: Applies to any relative pricing or speed claim. Does not apply to mechanism statements such as a cached-input discount, which can be checked on the pricing page. Breaks if the vendor publishes the full price table alongside the multiple. **Links**: - [[Reading vendor benchmarks]] — because extends - [[A precise number without a method is a marketing claim]] — because evidence-for the same pattern in the Jev video - [[AI cost architecture]] — because applies-in agent loops with cached context  #### [P2] Closed-set decision APIs are becoming a product category **Claim** (one sentence, Courtney's voice): When two vendors ship models that pick from a predefined option set, the interesting question is calibration and out-of-set behavior, not speed. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds when the answer set is enumerable and stable. The decisions API is in preview with no published semantics, so the claim is provisional until docs exist. Breaks if the API turns out to be a prompt wrapper over a general model. **Links**: - [[Classify before you generate]] — because extends - [[Jev Is Not an LLM (And That's Why It Wins)]] — because contrasts on vendor and claimed mechanism - [[Support ticket triage]] — because applies-in closed-set routing  ### ✏️ Fleeting Queue - [F1] Confirm the real model names (Astra, GPT-6.1 Sol, Ultrafast, Luna) and dollar prices from OpenAI's pricing page, since captions render "Astro," "Soul" and "ultraast" https://www.youtube.com/watch?v=Fls_onRviPM #fleeting #inbox - [F2] Find the docs for the decisions API preview and check out-of-set handling and confidence output before comparing it to Jev #fleeting #inbox - [F3] Identify the partner the caption renders "B 10" and the marketplace launch partners "burcell" and "Code Rabbit" from OpenAI's marketplace page #fleeting #inbox - [F4] Open question: how do dots that follow user access and act 24/7 interact with prompt injection through connected apps? #fleeting #inbox - [F5] Article seed: "Four stalled demos and no dollar prices" on how to read a keynote where the failures are disclosed and the numbers are not #fleeting #inbox #articleseed - [F6] Compare Codex Security's public beta scan findings with existing Tide and plugin-scanner output on one labeled repo #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Pull OpenAI pricing and docs for Sol, Ultrafast and the decisions API and record dollar prices and limits — research, 30 min #things - [ ] Draft a launch-day "verify before you cite" checklist for DevRel that covers relative multiples, caption garbles and coming-soon dates — DevRel enablement #things  📄 **Full text:** [[Full Document Contents/Articles/OpenAI DevDay 2026 Keynote.md|OpenAI DevDay 2026 Keynote by Sam Altman, Romain Huet, Tejal Patwardhan, Holly Li]] ## Highlights - [00:19:14] wherever it makes sense for you. And third, distribution. We want people We want to help people discover what you build and connect you with new customers. So, first let's start with the model. We launched Astra a few weeks ago. It's the most capable and aligned model in the market, but people keep asking us for two things. Make it cheaper and make it faster. And we really didn't want to disappoint you. So, first, we are launching GPT 6.1 Soul. You don't even know what it is yet. We got to tell you why it's good first. Um, this is a powerful model. Uh, it has a great balance of cost and capability. It [00:19:58] gives you very near Astro level intelligence at a fifth of the price to help you code and build. It's extremely affordable to run. In fact, the cache input costs are a 95% discount on standard input. So this makes it much cheaper for agents to reuse context, iterate on solutions, and work through longer tasks. We think you will really love this model that is smarter than Astra in some ways and it will become your daily workhorse. So we're very excited to get 6.1 Soul launched today. But sometimes you just want the highest intelligence at an absurdly fast speed. So today we're launching ultra fast across the API, Chat GBT, and Codeex. And we think you are really going to ([View Highlight](https://read.readwise.io/read/01m3qecng6v01bfyhs58hfehye)) - Note: [H1] Sol pitches near-flagship intelligence at a fifth of the price - Tags: atom/literature/h1 - [00:20:46] love this and not want to go back. Fast already gives you twice the speed at twice the price of standard, but ultraast takes that to eight times the speed at six times the price of standard. An incredible 300 tokens per second. You know what? It's worth it. Um, let me show you actually. So this is the same task side by side. We gave both models the same prompt. create and launch a rocket with the dev day colors. On the left here, you have ultra fast. On the right, you have regular. You can see how much faster this builds. The Ultraast rocket is already taken off and the regular rocket is not even close to being finished. There is nothing like this on the market today. You can use ultraast now with Astra and soon we're going to offer it for 6.1 soul as well. [00:21:31] So, here's what these improvements mean for your pro subscription. We're introducing Pro 500 with our highest usage limits and access to ultraast in chatbt and codecs. Pro 500 will have 25 times the usage of plus and you can use it across partner apps with signin with chatbt which we'll talk about later. We are also reopening the pro 200 subscription today. Pro 200 will continue to have access to all of our Frontier models. And with GPT 6.1 Soul, you will have a model close to Astro quality that you can use a lot as your daily driver. Our goal is to give you the best combination of intelligence and capability at every price point in the whole market. One more exciting speed update. Today ([View Highlight](https://read.readwise.io/read/01m3qecqykq4er640j8kvkgtd3)) - Note: [H2] Ultrafast trades 6x price for 8x speed - Tags: atom/literature/h2 - [00:22:14] we're previewing our new decisions API. The decisions API lets the model respond. Yeah, the decisions API lets the model respond in a fraction of a second is so fast. So it really feels almost instant to interact with. You can see here it using a computer and how quickly it's moving through these steps. This is not sped up. This works by giving our Luna model a predefined set of options to choose from like routing a request, classifying image, deciding what agent should do next. And by focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections. We're very excited to see what you'll all build [00:22:56] with this. So, thank you. So, OpenAI now offers leading models across intelligence, price, speed, voice, images, frontier models in every category. We want to be a one-stop shop where you can build. I'm really proud of our research team for getting the models to be this good, usable at this kind of scale, available at these price points, into the hands of developers everywhere. Now, there's a question of how the researchers can do all of this. Last year around this time, Yakob and I made a somewhat bold prediction on a live stream saying that we would see the first AI research intern in the next ([View Highlight](https://read.readwise.io/read/01m3qectesqay15rmjr8ykb5fp)) - Note: [H3] The decisions API restricts a model to predefined options - Tags: atom/literature/h3 - [00:26:32] longer. But as of July and with that good good ultra fast, our models have taken on over a third of dayong research task with no intervention required. So what does this mean? The mission of open AI is to ensure that AGI benefits all of humanity. And we want our models to help tackle some of the most complex challenges we are facing. Models like Astra have already helped solve over a 100 math problems that were open for decades. And they're also helping accelerate creating new antibiotics for drugresistant infections. studying ancient languages in history, developing better renewable energy and improving energy efficiency and accelerating industrial robotics and manufacturing. We believe our models will help make developers and users make very strong [00:27:14] progress on some of the most important and ambitious problems for humanity. And we're excited to see what you discover and build next. Thank you. And back to you, Sam. >> Thanks, Tachel. That was awesome. um we have a lot more stuff. So that brings us to the second part of supporting developers, giving you the tools to build what comes next. We want you to have access to the same foundations that we have at OpenAI. This starts with the codeex harness. It powers codecs work and now dots. We have put a lot of effort into making this harness very efficient so that you get accurate results at the fastest speed with the fewest tokens possible. And it's open source so you can look under ([View Highlight](https://read.readwise.io/read/01m3qed1r4s5xpzym9grkj6knm)) - Note: [H4] Research models finished a third of day-long tasks unaided - Tags: atom/literature/h3 - [00:47:03] going to simplify that. You'll be able to track review, see what needs fixing, request a human, and update your plugins tools without starting the whole submission over. So, finally, with all of this together, we want to make it easier to sell your product to more businesses. So today we're launching the OpenAI marketplace. We have 30 plus launch partners today including Code Rabbit, notion, burcell, and more. And our enterprise customers can now use part of their existing OpenAI commitment to buy your products. This gives you another way to reach customers already working with us. And of course, if you have an OpenAI commitment yourself, you can spend it in the marketplace, too. Through our partnership with B 10, you can even access open source models in the marketplace starting today. So thank you. [00:47:53] We shipped a lot of stuff this morning. New ways to work with AI with dots in space, new models with GPT 6.1 Soul and Ultraast, the decisions API, the agents API, codecs in the cloud, lots of new ways to help you build a business, drive consumers to you, and a bunch of other things. It's a big day for us and we're excited to see what you're going to do with all of this. But we wanted to end with what makes us so optimistic about this next phase of AI and that's all of you. Every day we hear stories about how you're using this technology to do things that last year would have felt out of reach. Working to understand the immune system, making education more accessible, ([View Highlight](https://read.readwise.io/read/01m3qed4e87eh1yh6dnrk7caka)) - Note: [H5] Marketplace lets enterprises pay with existing commitments - Tags: atom/literature/h3