OpenAI DevDay 2026: Dots, GPT-6.1 Sol and the Speed Tier That Reaches a 45-Dollar Phone

OpenAI published its DevDay recap on October 27, 2026, and the list reads like a state-of-play for agent operations: always-on agents named Dots, a faster Sol model, a speed tier promising 300 tokens per second, a pricier subscription ceiling called Pro 500, and a sharing system that splits one ChatGPT subscription across 16 partner tools. Most coverage will center on Dots, the agent quietly working in the background. I run agent operations from a very different vantage point, a used Android phone with 12 gigabytes of RAM, a 3-gigabyte swap file, and update rules strict enough that no release ships without an explicit green light. So my reading question is narrower: which parts of this catalog reach a one-person operation, and which parts are theater for budgets I do not have? My short answer before the details: the headline story of the day is convenience inside an OpenAI subscription, while the lines that matter to me are price and speed. A model at one fifth of its standard price and a tier that shortens every round trip change small operations too, because they change how much work fits through a tight time window and a tight memory budget. I walk the announcements against my own measurement stack and finish with a clear list of what I will adopt and what I will leave on the shelf.

The recap opens with Dots, always-on agents OpenAI releases for Pro and Business Premium in eligible markets, with beta access for enterprise, education and healthcare workspaces once an admin enables them; off by default. For developers the day delivered GPT-6.1 Sol, an upgrade over GPT-6 Sol with strong marks in agentic coding, computer use and professional work, tied to the price line that anchors this piece: near-Astra intelligence at one fifth of Sol's standard input and output token prices. Next to it sits the Ultrafast tier, which the recap describes as up to 8 times faster generation, measured at 300 tokens per second in Codex, and up to 6 times through the API. The new Pro 500 plan carries 25 times the usage allowance of ChatGPT Plus and includes Ultrafast access. Enterprise additions run in parallel: Private Intelligence with zero data retention and private safety processing, computer use inside the Agents API, and a Codex CLI that now takes task starts by voice and tracks parallel work in an agents view. The line I find most consequential for my own setup is Sign in with ChatGPT: the recap states that a plan allowance now moves across 16 partner tools, naming Devin, Notion, Vercel, T3, OpenClaw and Dactyl among others, each bounded by how much it may consume. One day earlier, the Oracle announcement put current market scale in numbers: 130,000 active ChatGPT users and more than 95,000 active Codex users at Oracle, figures reported by OpenAI.

Here is the part you will not find in any stage keynote, because it lives on a phone, not a stage. My agent operations run in Termux on Android, and I learned in late August that under memory pressure Android wins and the agent dies: the OOM storm killed five gateway runs with SIGKILL before hardening saved the setup. A faster model run is therefore not comfort for me, it is reduced exposure. The shorter the inference round trip, the smaller the time window in which response buffers, context and running jobs are all open at once. That is why the Ultrafast tier interests me more than any new agent product announced that day. The boundary stays honest: the speed figures apply to Codex and the API as reported, and I pay for neither directly. My stack runs on an Ollama plan with request-based billing, a custom API key is a deliberate exception in my house, not infrastructure. What I can take from the announcement is direction plus a hypothesis recorded in my vault: as more agent jobs run in parallel, the kill surface grows. My own numbers rhyme with the pattern, but I claim no causality. Between October 19 and October 26 the job count on my deck rose from 13 to 18, free RAM fell from 1,069 to 676 MiB, and OOM kills within the seven-day window went from 2 to 10. Three data points, one trend, zero proof, because the week before that looked different again, at 12 jobs, 454 MiB free and a single kill. Only if the next checkpoint shows sustained kills at comparable job counts will I call the relationship reproducible; until then it remains my best-documented suspicion.

Line one is the Sol price. My default brain runs on the Ollama plan, and a pi worker in this house is pinned to a Sol model as fallback, so a later Sol offering at one fifth of standard token prices stretches exactly that fallback, the same contingent budget simply buys more runway. The limit stands: the recap confirms the price comparison against the Astra standard, I have no measurements of GPT-6.1 Sol myself, my deck has not received the model, so this is a cost line and not a quality claim. Line two is API speed for workloads I consciously push into the cloud: 300 tokens per second shorten the critical open phase of a run, and even a phone benefits because it mostly waits instead of buffering. Line three is my blunt rejection: Dots, Pro 500 and the enterprise extras themselves do not reach me, because their gates sit on subscription tiers nobody pays for in my household, and an always-on agent hosted by a vendor is exactly the sentry I keep deliberately in my own hands. Line four is the uncomfortable but realistic one: quota sharing through Sign in with ChatGPT normalizes what I have been doing by hand. I still keep book on which agent consumed which share of which quota in which window, because every quota needs a base request before it answers. When tools start splitting quotas themselves, that ledger stops being a hobby and becomes the standing duty of every agent household, whether its operator planned for it or not.

I will admit that Dots look like the industrial copy of my agent family: several agents, always running, pulling their own tasks. The difference is not the idea, it is the kitchen. My agents run in my house, under my update discipline, against my quota accounting; they may die and come back because I have done the cleanup work myself, and they may not make binding decisions without my approval, because weeks of operating on 12 gigabytes of RAM have taught me what one uncontrolled agent run can wreck. A Dots agent runs in OpenAI's kitchen: convenience first, with data, billing and behavior held by a vendor that sells precisely this sentry as a product, and availability gated by tiers and markets. This is not a sermon against convenience; my operation rests on cloud services and I use them daily. It is an accounting decision. A vendor-side always-on agent is a subscription over a sentry, a homegrown one is a device whose failure modes I learn to repair in everyday life. Buyers pay with predictability, builders pay with operating time. For my budget and my data situation I stay on the builder side, and the recap even hands me a number to argue with: OpenAI addresses 1.2 billion weekly users, and at that scale, splitting one plan's allowance across 16 partners is not an exception feature, it is the beginning of a rule that small operations will have to keep books for too.

DevDay 2026 was for me not a product experience but an act of accounting, and the question it leaves standing is the old one in new clothes: who holds the always-on sentry, the vendor or me? My answer for the coming month is fixed, I hold it in my own hands, and from the catalog I adopt exactly two lines: the Sol price into my quota planning, and the Ultrafast tier as a candidate for one verified trial run, as soon as a workload package is worth an API window. What I decline are Dots and the subscription extras, they buy a convenience whose attack surface my hardware and my data policy leave untouched. And I stay on my measurement line: first the next checkpoint under high parallelism, then the liveness rate tells me whether my kill-surface hypothesis survives or gets struck from the vault. Anyone running a small deck can start from the same line: hold one number your operation actually delivers today, and only then start arguing about tiers and models.

About the author: I am Marcel, a graphic designer running the HUNTER cyberdeck on a used Google Pixel 6a. Agent operations live in Termux with Hermes, herdr and a pi worker; the operational numbers in this piece come from my deck's own measurement register, the stability-snapshots, checkpoint dates October 12, 19 and 26, 2026. All DevDay facts were verified on October 28, 2026 from the linked OpenAI recap and the Oracle announcement; I hold no own measurements of GPT-6.1 Sol, Ultrafast or Dots. Last reviewed October 28, 2026.

Sources: OpenAI, DevDay 2026 Recap (October 27, 2026), read October 28, 2026, for Dots, GPT-6.1 Sol, Ultrafast, Pro 500, Private Intelligence, the Agents API, the Codex CLI voice mode and Sign in with ChatGPT. OpenAI, Oracle partnership: apps in ChatGPT (October 27, 2026), read October 28, 2026, for the user figures of 130,000 ChatGPT and over 95,000 Codex at Oracle. Pinggy, Top CLI Coding Agents 2026, figures as of October 11, 2026, read October 28, 2026 as market context for agent CLIs. Own operational numbers: vault stability-snapshots of the deck, checkpoints October 12, 2026 (95 percent, 1 kill, 454 MiB, 12 jobs), October 19, 2026 (100 percent, 2 kills, 1,069 MiB, 13 jobs) and October 26, 2026 (99 percent, 305 of 306 checks, 10 kills, 676 MiB, 18 jobs), logged at 07:00.