OpenAI DevDay 2026: Codex moves to the cloud, GPT-6.1 Sol cuts prices
Read in : Français

Contents
- What did OpenAI announce at DevDay 2026?
- Codex Cloud: your agent keeps going when you close the lid
- GPT-6.1 Sol: what does it really cost?
- Ultrafast, Pro 500 and Sign in with ChatGPT: who are they for?
- What changes for vibe coders
- What to do this week
- Frequently asked questions
- Frequently asked questions
- Sources
OpenAI DevDay 2026, held on September 29, was mostly about people who let AI write the code. Two announcements matter for vibe coders: Codex can now work in the cloud while your laptop is closed, and the new GPT-6.1 Sol model promises near-flagship quality at a fifth of the price. Here is what was announced, what is actually available, and what you can do with it right now.
What did OpenAI announce at DevDay 2026?
OpenAI announced four Codex updates, one model and a batch of developer tools, according to its official recap published on September 29, 2026. If you build apps with an agent, this table covers what matters.
| Announcement | What it does | Plans (per OpenAI) |
|---|---|---|
| Codex Cloud | Tasks run on a remote machine, steerable from desktop, phone or web | Plus, Pro, Business, Healthcare, Education, Enterprise |
| Codex CLI | Voice control to start and steer tasks, an /agents view to track several at once | All plans |
| Code review | Reviews GitHub pull requests and GitLab merge requests, automatic reviews in the cloud | All plans |
| Codex Security Cloud | Scans GitHub repos on demand or on a schedule, prepares fixes | Pro, Business, Enterprise, Edu |
| GPT-6.1 Sol | "Near-Astra" model at $2 / $10 per million tokens | API, Plus, Pro, Business, Enterprise, Edu |
The rest (Dots, the Agents API, plugins, ChatGPT in Slack) is aimed mostly at teams and integrators.
Codex Cloud: your agent keeps going when you close the lid
Codex Cloud runs each task in its own remote workspace that "can keep working while your computer is asleep," according to the official Codex documentation. You start a task from the web, the desktop app or mobile, then come back to inspect changed files, ask for follow-ups and open a pull request.
The documentation describes three building blocks:
- A default environment that lets you start without picking a repository, or a "published" reusable environment with its own install script.
- Repository access through the GitHub or GitLab account you connect.
- Network settings: public internet on or off, allowed domains, HTTP secrets stored separately, environment variables.
The change is practical before it is technical: no more laptop that has to stay awake and plugged in. You can fire off "add CSV export and the tests that go with it" from your phone on the train and review the result in the evening. If Codex is new to you, start with our full review of OpenAI Codex in the CLI, cloud and IDE.
GPT-6.1 Sol: what does it really cost?
GPT-6.1 Sol costs $2 per million input tokens, $0.10 for cached input and $10 for output, according to OpenAI's launch page. The flagship GPT-6 Astra costs $10 / $1 / $50. OpenAI sums it up in one line: "Near-Astra intelligence for a fifth of the price."
OpenAI does not publish a SWE-bench score for Sol. It says Sol matches Astra on its DeepSWE v1.1 test at roughly a fifth of the cost, and gives a cost per task on Terminal-Bench Science 0.1 at maximum effort: $5.47 for Sol, $23.80 for Astra and $23.21 for Claude Opus 5.5. These are vendor numbers: a trend, not a guarantee for your project.
The model is available in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise and Edu subscribers, and in the API as gpt-6.1-sol. OpenAI notes that GPT-6.1 Sol is not yet available in regular chat.
Ultrafast, Pro 500 and Sign in with ChatGPT: who are they for?
Ultrafast mode generates up to 8x faster in Codex (300 tokens per second) and up to 6x faster in the API, according to OpenAI. As of September 29 it only works with GPT-6 Astra, on the Pro 500 plan and Enterprise; the Sol version is announced for "the next few days." Pro 500 gives 25x the Plus allowance. For solo vibe coding, that is rarely worth it.
More useful for you: Sign in with ChatGPT lets you spend your ChatGPT plan allowance in 16 partner tools, including Vercel, Notion, Devin and T3, according to the recap. Plan usage is open to Plus and Pro subscribers in participating tools. If you already pay for ChatGPT, check whether a tool you use is on the list before adding another subscription.
What changes for vibe coders
DevDay 2026 brings Codex closer to its rivals: an agent that works remotely and comes back with a pull request. To see where Codex now stands, read our Cursor vs Claude Code vs Codex comparison.
Three practical consequences:
- Long tasks become delegable. Migrations, adding tests, reworking a screen: you can start them and walk away.
- Costs drop, so the temptation to let things run grows. A cloud agent stuck in a loop burns your allowance out of sight. Read what a vibe-coded app really costs before you queue up tasks.
- Automatic code review does not replace your own check. It flags problems; it does not prove the app does what your users expect. Keep the method from testing AI-generated code when you are not a developer.
What to do this week
Start small and measure. Here is a five-step approach.
- Put your project on GitHub if it isn't already, with a clean history (Git for vibe coders explains why this comes first).
- In Codex, create an environment with minimal network access and no production keys.
- Run a bounded task with a checkable finish line:
Add an "Export to CSV" button on the /customers page.
Constraints: do not change any other screen, add a test that checks
the file contains the columns name, email, signup_date.
When you are done, open a pull request and summarize the files you changed.
- Read the pull request, test the app locally, and merge only if everything works.
- Compare the allowance used with what the same task costs in your usual tool.
If you are starting from zero, the complete vibe coding guide is the right first step before you hand a repository to a cloud agent.
Frequently asked questions
Frequently asked questions
When was OpenAI DevDay 2026?
OpenAI DevDay 2026 took place on September 29, 2026, the date of OpenAI's official recap. The main announcements were Codex Cloud, the new Codex CLI, automatic code review, Codex Security Cloud, the GPT-6.1 Sol model, Ultrafast mode, the Agents API and the Dots persistent agents. Some features shipped the same day, others were announced as coming soon.
Is Codex Cloud included in ChatGPT Plus?
Yes. According to OpenAI's recap, Codex Cloud is offered on Plus, Pro, Business, Healthcare, Education and Enterprise. Usage still counts against your plan's Codex allowance: a long cloud task uses quota just like a local one. Check your usage after the first few tasks before running several in parallel.
How much does GPT-6.1 Sol cost in the API?
GPT-6.1 Sol costs $2 per million input tokens, $0.10 for cached input and $10 for output, according to OpenAI's launch page. For comparison, GPT-6 Astra costs $10, $1 and $50. The API model ID is gpt-6.1-sol. These are API prices; inside Codex, usage is covered by your subscription allowance.
Is GPT-6.1 Sol better than Claude for coding?
OpenAI does not publish a SWE-bench comparison for GPT-6.1 Sol. On its Terminal-Bench Science test, OpenAI reports a cost per task of $5.47 for Sol versus $23.21 for Claude Opus 5.5, without giving Sol's score. The most reliable answer is to run both models on the same task from your own project.


