"Keeps Running Even When You Close Your Laptop" — At DevDay 2026, Codex Becomes an Always-On Development Platform and GPT-6.1 Sol Slashes the Price Tag to a Fifth

At "DevDay 2026," the annual developer conference OpenAI held on September 30 Japan time, more than 20 announcements were lined up. At the center of it all was a series of updates that reshape the coding agent "Codex" from a standalone command-line tool into a development platform that keeps running across your devices. Rolled out alongside it, the new model "GPT-6.1 Sol" and the high-speed tier "Ultrafast" form the groundwork for running that platform cheaply and quickly. In this article, we pull out and organize only the parts that matter on the ground for developers.

A Workspace That Doesn't Stop When You Step Away From Your Device

Codex can now save a development environment that bundles dependencies, settings, and permissions as a "reusable cloud environment." The key points are as follows.

• Continues across devices: A task you start on your PC keeps running on the cloud side even after you close your laptop, and you can check on it and operate it from another device such as your smartphone.
• Shared across the team: An environment with standardized settings and permissions can be shared across the team, and it is available on plans such as Plus / Pro / Business / Enterprise / Edu.

The premise of "first getting a repository set up locally, building the environment, and then working" fades, and the direction is one where the work state itself lives in the cloud.

The CLI Moves Toward "Directing by Voice" and "Bundling Multiple Agents"

Codex CLI, used day to day, was also overhauled. It now supports inserting instructions by voice, and along with that, the /agents view lets you assign work to multiple agents and track their progress in a single list. Running tasks like bug fixes, adding tests, and updating documentation in parallel is now opened up to all plans as a standard feature.

Turning Review and Security Into an "Automated First Pass"

Steps where a machine runs a first pass before a human looks have also increased.

• Code review: A new feature in the ChatGPT desktop app runs an "automated first review" on changes in the cloud and automatically returns those comments to GitHub pull requests or GitLab merge requests.
• Codex Security Cloud: It scans GitHub repositories on an "on-demand / scheduled" basis, scrutinizes the findings, removes duplicates, and even prepares fix proposals. This one is for Pro / Business / Enterprise / Edu.

Review and vulnerability response — steps that have carried a heavy manual burden until now — are positioned so agents slot in a first draft.

GPT-6.1 Sol, Which Cut the Price Tag to a Fifth, and Ultrafast at 300 Tokens Per Second

What underpins all this is the new model "GPT-6.1 Sol." It is said to deliver performance approaching that of the higher-end model "GPT-6 Astra" on agentic coding, computer operation, and specialized tasks, at roughly one-fifth the standard token price. On the tough software-engineering benchmark "DeepSWE v1.1," it is explained to achieve accuracy on par with Astra at about one-fifth the cost, exceeding the best score of the older "GPT-6 Sol" by 6.4 points. It is available via the API as gpt-6.1-sol starting the same day.

The published pricing (per million tokens, API) is as follows. It has been reconstructed based on the sources.

Cached input is said to be about 95% cheaper than standard input, and it starts to pay off on cost in iteration-heavy agent work. On the speed front, the newly established high-speed tier "Ultrafast" claims generation speeds of up to 8x on Codex (about 300 tokens per second) and up to 6x via the API. The shorter the wait, the more realistic it becomes to hand work off to agents and leave them running.

Also on the same day, OpenAI announced "dots," which stay resident in ChatGPT and work around the clock toward a goal, and "Plugin Extensions," which embed features into ChatGPT and Codex — strengthening, overall, the push toward keeping agents constantly running.

Behind Cheaper, Faster, Self-Running: Where Do You Put the Eyes That Inspect?

Development speed and unit cost do indeed come down. At the same time, as agents that stay resident and run on their own — and mechanisms that go so far as to "find and fix" vulnerabilities — spread, the question of who reviews those judgments remains. Automated review and automated fixes are handy as drafts, but if they are treated as the final decision, there is a concern that errors introduced by a machine could slip past human review unchecked. And many of the benchmarks are figures published by the provider itself; independent third-party verification is still to come. The more the competition over price cuts and speed continues, the more it becomes necessary to weave into operational design — alongside cost-effectiveness — the question of "where to keep human inspection."

That said, the combination of an environment that keeps running unbound from your device, parallel work bundled by voice and /agents, and an automated first pass for review and security is a practical change for both individual developers and teams. A realistic way to start is to try automated review and security scans on a limited repository first, and decide where to draw the boundary that keeps human review in place.

References: ITmedia NEWS, "A Summary of the Main Things Announced at OpenAI's 'DevDay 2026'" / Unite.AI, "OpenAI Unveils GPT-6.1 Sol at DevDay With New Codex and ChatGPT Tools" / The Next Web, "OpenAI releases GPT-6.1 Sol at a fifth of GPT-6 Astra's token prices" / 9to5Mac, "OpenAI makes 20+ announcements at DevDay event including always-on agents called dots"