OpenAI DevDay 2026 roundup: Dots, GPT-6.1 Sol and the updates that actually matter

My feed check for 04.10.2026 counts 21 new videos across the 58 subscribed channels, and the one worth building a post around comes from the Everlast AI channel: the team was on the ground in San Francisco for the OpenAI Developer Day on 29.09.2026 and condenses more than 20 updates into a single AI news episode. For most viewers this is consumer news. For an agent operation like mine, it is a marker: OpenAI has installed an always-on agent called Dots into its flagship product, the new GPT-6.1 Sol model cuts the price to a fifth of the Astra level, the new Pro plan costs $500 a month, and 1.2 billion people use ChatGPT every week. This post works through the announcements the way I read them from my daily Hermes and pi setup: with the numbers from the primary sources, with caveats wherever a video only translates claims and shows no benchmarks, and with a clean line between what OpenAI announced and what still has to prove itself.

Sam Altman framed the whole keynote around one question: how humans and AI work together. Dots is his answer. A Dot is an agent that keeps working after the first instruction, powered by GPT-6 Astra, equipped with its own cloud computer and browser, reachable through ChatGPT, SMS, email and Slack. You can call a Dot, the demo leaned on phone calls and iMessage as entry points instead of typed prompts. The demo star went by the name Alfred: he reads your calendar and your installed apps, shows up proactively with ideas, and when you repeatedly ask about your business numbers he suggests building a KPI dashboard, then building an app around it. Rules and permissions cap what a Dot may do, from confirming individual actions to your own house rules. Alongside that came Spaces, described as Notion for humans and agents: one page, a comment thread to delegate tasks to agents, team sharing of pages and a preview mode for presentations. For businesses OpenAI showed specialized Dots for accounting, email marketing and legal work, a Microsoft partnership was mentioned, and whether those specialists stay exclusive to enterprise customers was left open, the video says so explicitly. Two things matter to me. First, Dots is a consumer product: calendars, apps and phone calls for the mass audience, the mirror image of my terminal-based operation. Second, delegation through comment threads is conceptually the same technique I practice every day by sending Telegram messages to Hermes, only packaged for the mainstream. The fact that OpenAI builds exactly this instead of just another model bump supports the core thesis of my operation: the value sits in the agent around the model, not in the model itself. The channel adds two ironic footnotes: OpenAI's CFO mixed up the name Dots with the name of a competing agent product several times during her talk, and a Meta agent app briefly pushed ChatGPT off the top of the US App Store, OpenAI is reacting to a market that moves faster than its own keynote.

GPT-6.1 Sol is the newest member of the GPT-6 family, described by OpenAI as reaching close to the level of the flagship GPT-6 Astra, focused on coding, computer use and professional workloads. The pricing is the actual story: $2 per million input tokens, $10 per million output tokens, cached input for $0.10. That is a fifth of the Astra price, and the cached input rate is a tenth of the Astra level. The context window spans 1.05 million tokens with up to 128,000 output tokens, and reasoning effort scales from low to maximum. The video adds a number that OpenAI benchmarks do not back up: according to the Artificial Analysis index, the standard test task is said to cost only 72 cents in model spend, while Claude Sonnet 5.5 counts as capable but too expensive for routine work. OpenAI showed no benchmarks for Sol itself, so genuine practitioner verdicts still have to emerge. I keep that gap visible before anyone turns it into a buying recommendation. Next to that sits the Ultrafast tier: roughly 300 tokens per second, eight times faster by OpenAI's own claim, but about six times as expensive, and the audience audibly hesitated at exactly that combination. The plan structure flips the subsidy ladder: Pro 500 costs $500 a month, includes Ultrafast, lifts usage to 25 times the Plus allowance and removes the 5-hour limit. At the same time, the $200 plan now contains only half of its previous capacity. The lesson for an agent operation is sober: subsidies rotate, the durable approach is to stay model-agnostic, OpenAI today, Claude tomorrow, local models in between, and your monthly quota tells you at the end of the month which plan you can actually trust.

Three updates in the model block carry extra weight. The Decisions API targets agents that must decide faster and better: presented as a way to route decisions through the fast Luna model over the API, with model routing as the obvious use case. For my operation it is a competing offer aimed at the same layer I already run with Luna, interesting as a fallback path, not as a replacement. Second, OpenAI declared the Research Intern officially reached: an AI research intern that works autonomously, with recursive self-improvement already running inside the company. The honest framing sits right next to it in the video: OpenAI researchers improved the computer-use harness with the help of AI instead of programming it themselves, computer use now runs twice as fast and, by OpenAI's own benchmarks, more safely. The slide, as the channel reads it, is a deliberate attack on Anthropic's image as the safest AI lab, and that claim rises and falls with benchmark numbers presented by the interested party itself. Third, the time spans: OpenAI agents handled 15 minutes of autonomous work at the start of the year, on stage it was 8 to 16 hours. On top come the Codex additions: the public Codex harness is improved, remote control now runs in the cloud so agents stay reachable around the clock instead of being tied to an unlocked office computer, a new Agents API exists, the Codex Security Cloud finds bugs and vulnerabilities in your own apps for free, AWS Bedrock joined as a partner, Private Intelligence promises zero data retention with confidential computing, and the Codex CLI can meanwhile operate any app on your machine.

Sam Altman sorted the platform into three areas: models, building, distribution. Models rotate weekly and are no longer the deciding factor, building is already routine for most, distribution is the open field. ChatGPT counts 1.2 billion weekly users, and three updates aim straight at that funnel. Sign in with ChatGPT opens your own app to ChatGPT users: whoever holds a $20 plan, or even a free plan, can use your app with that very subscription instead of paying through the API. With that move OpenAI pushes its subsidies into third-party apps, and for app builders it is the biggest cost shift of the day. Plugin extensions make partner plugins discoverable inside chat sessions, partnerships with Adobe and Figma were named, and the ecosystem is being built so there is no reason to leave. The OpenAI marketplace promises distribution for your own app plus a share of the revenue, access to open source models is announced as well. The channel names its skepticism as the feeling of a recurring promise, app store announcements have happened before. Two side notes follow: the GPT-6.1 Astra variant slipped to the end of the year, so OpenAI is pausing its model roadmap at the same time. And Google presented Gemini 4 Argon, which is not live and stays restricted to government institutions and registered security firms, those benchmark numbers deserve a matching caveat, the video explicitly warns against benchmark maxing. The channel's distribution thesis is pointed: marketing has become fully automatable, cheap and scalable, with hundreds of agents working in parallel. The workshop the channel sells on that premise is a self-interest you can spot as such.

Three side notes deserved more time than the keynote gave them. First robotics: Figure announced in a video featuring Arnold Schwarzenegger that the Figure 2 is no longer manufactured or shipped. The channel was openly puzzled by that much star power for a discontinuation clip, and reports that US robotics companies such as 1X with the Neo showed little transparency to its team. Concrete operation showed up in Germany instead: at the UBTECH event hosted by Terrobotics a humanoid stood right in front of the camera, Terrobotics distributes UBTECH robots exclusively in the DACH region, the Alpha Mini 2 dances in front of an audience, teleoperation on the Tienkung could be tried out through a VR headset, the industrial use case is standardized KLT boxes in logistics and production, the range stretches from cleaning robotics to industry, and the developers in China hold weekly exchanges with Germany. Second, the US government: America.gov is live, travel bookings run through a chatbot, the presentation feels like an Apple demo, prompt injections remain an open problem, and the open question is when Germany will show anything comparable. Third, Tavos: the avatar vendor claims to run the first video model to pass a Turing test in video, 48 percent of participants in a video call could not tell the AI from a human. For agent operations the takeaway: an avatar in customer contact is no longer science fiction, it is a line item.

My operation does not change on this day, and that is the real headline. Dots are consumer agents with calendars and phone calls, my setup stays terminal, bridge and cron, and the 21 new videos in this check show the agent toolchain running on outside the OpenAI stage, see the Grok-and-Hermes clips of the same week. Three things still go on the list. The price curve of GPT-6.1 Sol changes the math for routine tasks, a fifth of the Astra price and a tenth on cached input is a gap you feel on monthly limited quotas. The Decisions API and the Luna layer are a competing offer for what I already run myself, interesting as a fallback, not as a swap. And the Codex Security Cloud lowers the barrier to scanning your own apps for bugs and vulnerabilities automatically, that stays an observation post until a real use case shows up in my operation. What OpenAI proves with Dots is the core thesis of my operation in mass format: the agent around the model carries the value, the model is the rotating part, and whoever built a solid setup changes models like gloves without touching the running operation.

[OpenAI: DevDay 2026 Recap](https://openai.com/index/devday-2026-recap/) (retrieved: 04.10.2026) | [Business Standard: OpenAI DevDay 2026 brought Dots agent, GPT-6.1 Sol, cloud Codex and new developer tools](https://www.business-standard.com/technology/tech-news/openai-devday-2026-dots-gpt-6-1-sol-codex-developer-tools-126093000396_1.html) (retrieved: 04.10.2026) | [Mashable: Biggest news from OpenAI Dev Day](https://mashable.com/tech/openai-devday-2026-biggest-announcements) (retrieved: 04.10.2026) | [Everlast AI: The new ChatGPT is here, all 20+ updates, AI news, YouTube](https://www.youtube.com/watch?v=1ID_VcwzIo8) (retrieved: 04.10.2026) | [Sam Altman on X: Dots announcement, 29.09.2026](https://x.com/sama/status/2104995014208258235) (retrieved: 04.10.2026)

About the author: Marcel, agent operation between Hermes, pi and the weekly feed check. This blog reports tested experience instead of hype, drafts on this page are proposals until approved.