OpenAI DevDay 2026 Recap: Dots, GPT-6.1 Sol, Pro 500 and Every Other Announcement

Everything OpenAI announced at DevDay 2026

The most consequential line in OpenAI's DevDay 2026 recap is not a model name. It is a description of what a dot does while you are not looking: it "looks for ways to help in the background," reading the apps you have connected, and brings you work "sometimes before you even think to ask." ChatGPT has always waited for a prompt. The product OpenAI put first on 29 September does not.

Around that sat a cheaper frontier-class model, a speed tier and a $500 plan to pay for it, and a set of platform changes that let developers put their own software inside ChatGPT and let users spend their ChatGPT plan inside other companies' apps. OpenAI grouped the launches in five parts. Here is all of it in one table, with the availability OpenAI stated.

The sections below take the launches in order of how much they change what developers and businesses should do next.

Dots: agents that keep working after you close the tab

A dot is an agent with its own cloud computer and its own browser, running on GPT-6 Astra. OpenAI's launch post says dots connect to more than 4,000 apps through ChatGPT's plugins, learn from feedback, and work toward your goals around the clock. You name your first one, message it in ChatGPT on desktop, web or mobile, in Slack or in Teams (texting is "coming soon"), or call it by voice. It can take on several projects at once, and you can open its computer at any time to see what it is doing.

The background work has limits that matter. OpenAI calls it "proactive research" and says it uses tools restricted to read-only: a dot looking for ways to help cannot send messages, change app content or control your computer while it does. Actions that touch accounts or share information pass through an auto-review against your instructions and Custom Rules, and some tasks, such as changing a password, always stay with you. Your own laptop stays separate unless you connect it.

Dots roll out to Pro and Business Premium users in eligible markets, and the first dot is included at no extra cost. Talking to your dot does not count against ChatGPT usage limits; tasks it starts in Codex or ChatGPT Work do. For Enterprise, Edu and Healthcare the beta is off until a workspace admin turns it on. OpenAI also previewed specialist dots, agents a company sets up with their own identity and credentials for a defined job, which start as enterprise pilots, and said it is working with Microsoft to manage them through Agent 365.

The example OpenAI chose to illustrate all this is modest on purpose: an early tester's dot noticed he had forgotten to invoice a publication, prepared the invoice and sent it after he approved. That is the pattern to expect from the first version. Small chores caught early, with a person approving anything that leaves the building.

GPT-6.1 Sol: near-Astra results at a fifth of the price

For most developers the money story is GPT-6.1 Sol. In the API, gpt-6.1-sol costs $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. OpenAI says that is one-fifth of GPT-6 Astra's standard input and output prices, and the cached rate is half what GPT-6 Sol charged. For agents that resend the same long context on every call, the cache price is the number to plan around.

OpenAI's own benchmark claims, as it stated them:

• DeepSWE v1.1 (software engineering in real codebases): matches GPT-6 Astra at roughly one-fifth of the cost, and beats GPT-6 Sol's best score by 6.4 points.
• AutomationBench (multi-step business workflows): 2.2 points above Claude Opus 5.5 at medium effort, at roughly a third of the cost, and 4.8 points above GPT-6 Sol.
• OSWorld 2.0, offline set (computer use): 7 points above GPT-6 Sol at maximum effort, and within 2.1 points of Astra at roughly one-seventh the cost per task.
• Terminal-Bench Science 0.1: $5.47 per task at maximum effort, against $23.21 for Opus 5.5 and $23.80 for Astra. Astra still posts the top score, 68.1%, and OpenAI says to use it for the hardest research.
• Factuality on deliberately hard prompts: responses with a factual error fell from 11.4% to 7.7% at low effort compared with GPT-6 Sol.

These are vendor numbers on vendor-chosen settings, and several compare cost per task rather than accuracy alone. Treat them as a reason to test, not a conclusion. The availability detail is the one people will trip over: in ChatGPT, GPT-6.1 Sol is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and "not yet available in Chat." An Ultrafast version is promised "in the coming days."

Ultrafast, and a $500 plan to pay for it

Ultrafast is OpenAI's premium speed tier: up to 8 times faster token generation in Codex, which OpenAI puts at 300 tokens per second, and up to 6 times faster in the API. GPT-6 Astra Ultrafast is live in the API and in ChatGPT Work and Codex on Pro 500 and Enterprise; GPT-6.1 Sol Ultrafast is "coming soon."

In the API it is a service tier, not a separate model. The documentation sets service_tier to ultrafast on gpt-6-astra, recommends WebSockets because per-request network overhead can eat the gain in tool-heavy agents, and starts every organization on low limits: 500,000 tokens per minute for usage tiers 1 to 3, 1 million for tier 4 and 5 million for tier 5. It supports US data residency and global processing only, so EU regional processing is out for now.

In ChatGPT, speed now has its own price point. OpenAI's Pro tiers page lists three monthly plans:

Pro 200 is open to new subscribers again, with a smaller allowance than before. Subscribers who were active at the eligibility cutoff, or in the seven days before it, keep their previous allowance through 29 October 2026, then move to the new one at the same $200. Buying credits on Pro 100 or Pro 200 does not unlock Ultrafast.

Agents API with computer use, the Decisions API and Bedrock

The Agents API, in public beta, packages the harness that runs Codex: context compaction for sessions that outgrow one context window, tool search that loads tool definitions only when needed, programmatic tool calling, and subagents that work in parallel. It supports MCP, custom functions and built-in tools such as web search. You choose where the agent runs: an OpenAI-hosted sandbox, your own infrastructure, or a partner sandbox from Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop or Vercel. There is no separate fee; you pay for tokens and tools. The harness is the open-source Codex code on GitHub, so you can read what it does.

DevDay's addition is computer use, so agents built on the API can operate software, not just call functions. It is available through the API and in Codex and ChatGPT Work on Pro 500 and Enterprise.

The quieter launch is the Decisions API. OpenAI describes it as focusing Luna's intelligence "on a specific set of user-defined questions with finite pre-defined answers." You send text or images and get back an answer you can use to classify content, route a request or pick an agent's next step. That is a different product shape from a chat model: a fixed menu of answers instead of free text, which is what most production pipelines actually need. It is in limited preview, with a broad release planned in the coming days. OpenAI did not publish a price in the recap. If you have been following decision models such as TypeSafe's Jev, this is OpenAI entering the same category.

For companies that live in AWS, Bedrock Managed Agents, built with Amazon, take the core of the Agents API and run OpenAI agents entirely inside AWS, wired to AWS resources.

Codex moves to the cloud, and into security

Codex now runs on your computer, remotely from a phone, or in the cloud from any device, with reusable development environments that give a team a shared setup and approved permissions. That reaches Plus, Pro, Business, Healthcare, Education and Enterprise. The CLI gained voice control for starting and steering tasks, an /agents view for tracking several tasks at once, and better handling of prompts, resumed sessions and worktrees.

Code review moves into the ChatGPT desktop app: summaries, diffs and questions to Codex before you comment on a GitHub pull request or GitLab merge request, plus automatic first-pass reviews in the cloud. Codex Security Cloud scans whole GitHub repositories on demand or on a schedule, checks new commits, removes duplicate findings and prepares fixes in the cloud with your laptop closed. It includes models offered through OpenAI's Daybreak Blue without a separate application, on Pro, Business, Enterprise and Edu.

How this stacks up against Claude Code, Cursor and Antigravity after this month's releases is in our AI coding tools comparison.

ChatGPT opens its own platform to developers

"We're opening the platform we use to build ChatGPT features," OpenAI wrote. Plugin extensions give a plugin a home in ChatGPT's sidebar, interactive panels that sit beside the conversation, and viewers for your product's file types, on all plans. A Plugin Creator helps build them, a redesigned submission flow gives clearer feedback, and OpenAI says improved ranking and recommendations will surface plugins "in the directory and in conversations," with users approving each plugin's access. Business, Enterprise, Healthcare and Edu workspaces can now put plugins inside Sites. And plugins can trigger automations when something happens in a connected app, through support for the proposed MCP Events specification.

The other half is money moving the other way. Sign in with ChatGPT lets anyone log in to participating apps, and lets Plus and Pro users spend their ChatGPT Work and Codex allowance inside 16 partner tools, including Devin, Notion, Vercel, T3, OpenClaw and Dactyl, with a weekly cap per app set in Settings, Usage. The app receives your name, email and profile picture. It does not get your conversations, your memories or an API key. Airtable, Canva, GitLab, HubSpot and Supabase support the sign-in but not plan sharing yet.

For a developer the combination is new: distribution inside a product with 1.2 billion weekly users, and a way for those users to pay for your AI features with a subscription they already have.

Spaces, Pages and work shared with agents

The team features all assume a dot or ChatGPT is a member of the team. ChatGPT Space is a shared home where teammates, ChatGPT and your dot build on the same knowledge, kept organized by instructions you give it (Pro, Business and Enterprise on desktop and web; reading and sharing on mobile, creation there later). Pages are documents built for people and agents to edit together, and can hold anything a conversation can: writing, research, charts, images. Collaborative slides arrive in the coming weeks, with several teammates and agents editing at once and export to PowerPoint or Google Slides.

On Business and Enterprise, teams share pages, slides, plugins and spreadsheets, and team tasks run recurring work on a schedule or on a trigger such as a new email or Slack message. @ChatGPT can be mentioned in Slack and Microsoft Teams channels, threads and DMs, and teammates can join in without their own license. The Meetings plugin, in beta on the macOS desktop app for Pro and Business, saves notes and action items to Space and deletes the audio once notes are ready. Shareable profiles let people publish the Sites and plugins they have built for others to reuse.

Private Intelligence and the OpenAI Marketplace

Two launches are aimed at the buyers who sign large contracts. Private Intelligence starts with Zero Data Retention with Private Safety Processing: automated safety reviews without OpenAI staff seeing the content. The setup guide has organizations already approved for Zero Data Retention connect their own AWS S3, Azure Blob or Google Cloud Storage bucket, enabled per project, with a 30-day lifecycle rule on the stored records. A preview of Private Inference, combining confidential computing with verifiable controls, is due this fall.

The OpenAI Marketplace lets eligible enterprises put part of an existing OpenAI commitment toward approved partner software. The first 32 partners include Adobe and Figma, Sierra, Decagon, HubSpot, Salesforce and ServiceNow, Harvey and Legora, Palo Alto Networks and CrowdStrike, and Baseten for open-source models. For those vendors, a customer's OpenAI budget just became a way to buy them.

What DevDay means for search and marketing

None of this was announced as a search product. Three parts of it will still change how people find businesses, and it is worth being clear that what follows is our reading, not OpenAI's.

First, ChatGPT now has a ranking system for software. OpenAI says improved ranking and recommendations will help people find plugins in the directory and inside conversations. It did not say how that ranking works. For any company with a plugin, that is a new place to be found or overlooked, with rules that are not yet documented.

Second, more of the looking will be done by agents. A dot doing proactive research through a browser and 4,000 connected apps is choosing sources and vendors on someone's behalf, often before the person asks. Whether AI assistants name your business stops being a vanity question when an agent acts on the answer. It can be measured, as a rate with a margin of error rather than a screenshot; our guide to measuring AI visibility sets out how, and it is what the AI search surface in AIO Copilot tracks for each client.

Third, classification got cheaper to build. Search teams label things all day: search terms as waste or not, pages by intent, queries by topic. A Decisions API that returns one answer from a fixed list is built for exactly that, and it keeps the arithmetic out of the model, which is where it belongs.

What OpenAI did not say

• The Decisions API's price, and the date of its broad release beyond "the coming days."
• Ultrafast's token prices were not in the recap; they sit on OpenAI's API pricing page, and GPT-6.1 Sol Ultrafast has no date.
• When GPT-6.1 Sol reaches ordinary Chat in ChatGPT.
• Which markets count as eligible for dots, beyond a help page, and when more dots or more users arrive.
• How plugin ranking and recommendation actually work.
• A firmer date for Private Inference than "this fall."

If you run agents on GPT-6 Sol today, the cheapest thing to do this week is test the same workload on gpt-6.1-sol with caching on, and measure the cost per finished task rather than per token. If you sell software, read the plugin extension and Sign in with ChatGPT docs before your competitors do. And if customers find you through AI answers, start measuring how often they do now, because the assistants doing the asking are about to multiply.

Frequently Asked Questions