OpenAI DevDay 2026 Puts Codex in the Cloud for AI-First Coding

Prompting is becoming delegating

At DevDay 2026, OpenAI did not lead with the model. The main story was what happens to software work once writing code by prompting becomes routine. At the September 29 event at Fort Mason in San Francisco, OpenAI shipped more than 20 announcements. The Codex coding agent was a big part of them. It can now keep working after a developer closes the laptop, review its own pull requests, scan repositories for security holes, and serve as the engine developers use to build their own agents91. Read together, the coverage shows a company trying to turn casual "vibe coding" into a continuous, agent-run development process. Most outlets agree on what shipped. They disagree on whether OpenAI can make all of it work together.

InfoQ's recap summed up the theme: OpenAI is moving from single model interactions toward persistent agent workflows, building infrastructure for agents that work across applications and tasks without stopping3. CometAPI called the overall shift "executable AI." It argued that Codex now looks more like a development environment that carries out tasks than a code assistant18. That is a fair reading of the announcements. The details show how far the plan has actually gotten.

Codex leaves the laptop

The most important change for everyday AI coding is Codex in the cloud. Developers set up reusable environments that bundle repositories, dependencies, scripts and access settings. Tasks keep running on managed machines after the developer's own computer is off, and progress can be checked or steered from another device9. One practitioner's walkthrough explains the mechanics: Codex inspects the repositories, prepares the setup, tests it with the user, and then publishes the environment. After that, each new task gets its own isolated workspace17. The feature covers Plus, Pro, Business, Healthcare, Education and Enterprise plans, but not Free or Go17.

The keynote demos showed how OpenAI thinks about prompt-driven app building. On stage, Codex built a 3D space game, and a presenter asked it to rewrite a flight-journal app he had vibe-coded so that it ran in Rust, leaving the job running in the background12. Simon Willison's live blog recorded a related demo of Codex Remote building an iPhone app, with the simulator's interface shown inside Codex11. A later session showed a prompt that turned a Google Sheet into a "launch radar" web app. Codex could then be told to watch the sheet and keep the published site up to date11.

That last demo matters because it changes what a prompt is. In early vibe coding, a prompt produced code once and the conversation ended. Here, a prompt sets up an ongoing duty. The rest of the DevDay announcements point the same way.

The terminal becomes a control room

The updated Codex CLI, available on all plans, lets developers start and steer tasks by voice. It also adds an /agents view for handing off and tracking several tasks at once, plus improvements to prompt editing, session resuming and Git worktrees194. One independent analysis argued that worktrees are the real advance, not voice. Two agents working in the same directory can overwrite each other's files, while separate worktrees give each agent its own branch to merge like a normal pull request. The same writer noted that voice does nothing to raise how much agent output a team can review in a day, and that review is usually the real limit17.

That review limit explains why OpenAI paired the CLI update with review tools. A new code-review view in the ChatGPT desktop app shows summaries and diffs across projects, sends feedback to GitHub pull requests or GitLab merge requests, and can do an automatic first pass in the cloud while the developer is away19. The coverage differs slightly on maturity. Several recaps describe review as live on all plans4, but one detailed breakdown says GitHub support is generally available while GitLab merge requests are still in preview17. The more careful description is probably the accurate one.

Security agents that propose but don't merge

Codex Security Cloud scans GitHub repositories on demand, on a schedule, or as new commits arrive. It investigates and removes duplicate findings, then prepares fixes313. One account describes a four-step process: build a threat model, scan, try to reproduce likely vulnerabilities in a sandbox to filter out false positives, and propose a patch for a human to review before a draft pull request is opened. OpenAI presents the tool as a complement to static analysis, not a replacement17.

In a breakout session, OpenAI staff said that patch generation was introduced gradually as trust grew. They reported only a 1% rollback rate, which they credited to a verification step that acts as an adversarial agent testing each fix. A human still decides whether to merge11. Release status varies by account here too. One recap lists the product as live4, while another calls it a research preview on Pro, Business, Enterprise and Edu plans17. For teams deciding whether to rely on it, that difference matters.

This is the clearest example of the AI-first development model OpenAI is promoting. Agents do the high-volume work of finding and fixing, and humans hold the approval gate. It is a sensible design, and it admits openly that generating code is no longer the bottleneck.

The harness goes on sale

For developers building their own AI products, the main news is that OpenAI is offering the machinery behind Codex. The Agents API, in public beta, now supports computer use, so agents can operate software through its graphical interface. It also brings Codex's multi-agent features, tool search, tool calling and context compaction, with OpenAI running the execution infrastructure23. One analysis says OpenAI manages the agent harness, durable sessions and orchestration, while developers supply the tools and decide where code runs. It concluded that much of the work of running long agent sessions is shifting from app developers to platforms9.

The new Decisions API, in limited preview, adds a narrower building block. It uses the small Luna model to pick from answers the developer defines in advance, which suits classifying, routing, or choosing an agent's next step219. Amazon's Bedrock Managed Agents service runs OpenAI-powered agents inside AWS2.

The models underneath are priced for agent-scale use. GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding and computer use by OpenAI's own evaluations, at one-fifth of Astra's standard token prices15. One recap puts it at $2 input and $10 output per million tokens5. Those benchmarks come from OpenAI and have not been independently verified4. Ultrafast, a premium speed tier, claims up to 8x faster generation in Codex and 6x in the API2. The Decoder reports it costs six times as much19, and outside Enterprise it is limited to the new Pro 500 plan5. CometAPI notes that faster token generation does not mean tasks finish proportionally sooner18.

In the closing Q&A, Codex product lead Thibault Sottiaux said Ultrafast helped OpenAI merge the ChatGPT and Codex desktop apps in 28 days11. That is OpenAI vouching for its own product, but it shows how much the company itself depends on agent-driven development.

Apps you build inside the chat

Plugin extensions bring app building to ChatGPT itself. Developers can build sidebar entries, interactive panels and file viewers, which Altman described as plugins that are "essentially entire applications" running inside ChatGPT and Codex12. Plugins can also react to events through the proposed MCP Events specification, which lets automations start when something happens in a connected app3. OpenAI is pitching this as access to ChatGPT's 1.2 billion weekly users4. Sign in with ChatGPT lets Plus and Pro users spend their plan allowance in 16 partner tools, including Devin and Vercel5, which changes who pays for AI usage in third-party coding products.

Altman also said directly how he hopes to reach builders who don't think of themselves as builders. He acknowledged that coding agents are much harder to get started with than ChatGPT. His hope is that people will talk to their Dot, OpenAI's new always-on agent, during the day, and the Dot will go off and build things11. In that picture, building an app from prompts disappears into ordinary conversation.

The reception: ambitious, uneven

The coverage agrees on the scope of the launches. It does not agree on how well they work. InfoQ and daily.dev both describe a mixed reaction from developers: some valued cloud Codex, computer use and cheaper models, while others doubted these went much beyond existing agent platforms1415. Engadget's live blog included reader comments saying Anthropic was ahead in coding and that DevDay needed to give developers something they would actually use12. Willison wrote that he tried to build his live-blog photo tool with Codex Cloud on his phone that morning, hit problems, and switched to Claude Code for web11.

Every's review, published three weeks after the event, made the sharpest criticism. It called the pace exhilarating and said manually managing individual Codex threads already feels outdated. It also warned that shipping so much at once risks losing the simplicity that made Codex good, and admitted that even a power user was tired, confused and dealing with rough edges13.

That critique seems right. DevDay 2026 sets out a coherent vision of AI-first software development: prompts start ongoing work, agents run in the cloud, review and security are automated but gated by humans, and finished apps ship straight to ChatGPT. The question for the coming months is whether the pieces fit together well enough that developers trust them with real work. The direction is clear, but the reliability has to be proven before working developers will commit to it.