OpenAI Unveils New Codex AI Agents and GPT-5.6 Tools for Software Developers
OpenAI has spent the back half of August rolling out a dense stretch of updates aimed squarely at developers and the engineering teams managing them, and taken together, the changes mark one of the more substantial shifts in how the company positions Codex, its AI coding agent, as a genuine workplace tool rather than just a chatbot with programming knowledge.
The centerpiece of the recent push is the new Admin plugin for ChatGPT Work and Codex, which OpenAI introduced to bring workspace analytics, access controls, and routine administrative actions directly into a single conversation. Rather than bouncing between separate dashboards to check usage, adjust permissions, or approve spending requests, admins can now ask a question, review the details, make an authorized change, and confirm the outcome without leaving the chat interface. Kunal Malik, OpenAI’s Head of Global IT, described the plugin’s value as extending well past faster reporting, since it connects a question directly to the next supported action, whether that’s checking someone’s permissions, adjusting a usage limit, or reviewing a spending request. Internally, OpenAI says a similar Slack-based agent has been resolving roughly 45 percent of employee IT ticket volume, a meaningful chunk of routine support work being handled without a human stepping in first.
Alongside the Admin plugin, OpenAI introduced Site tools built on WebMCP, a proposed open standard that lets websites expose specific actions directly to AI agents rather than forcing the agent to navigate a page the way a human would. In practice, this means a website can offer functionality straight to ChatGPT Work or Codex through a shared, signed-in session, letting the agent and the user work against the same live page simultaneously. OpenAI has framed this as a potential successor to Custom GPTs, since a business could expose its own data and tools through WebMCP and let a user’s existing ChatGPT or Codex session call into that functionality directly. The feature currently works with GPT-5.6 Sol and GPT-5.6 Terra in the desktop app’s built-in browser, though it isn’t yet available with the lighter GPT-5.6 Luna tier or within Enterprise and Edu workspaces.
Browser support for Codex and ChatGPT Work also widened this month. The ChatGPT browser extension, previously limited mostly to Chrome, now works across Microsoft Edge, Brave, Opera, and Vivaldi as well, with tab mentions and browser control supported across all five, though Opera lacks side chat functionality. OpenAI also added cloud browser sign-in for eligible plans, letting ChatGPT Work request website logins through a secure cloud-based browser session that stays separate from a user’s local browser, useful for agents that need to complete multi-step tasks on sites requiring authentication without exposing credentials to the model itself.
Underneath all of this sits a broader model transition that’s been reshaping Codex through most of the summer. GPT-5.6 became generally available in three tiers back in July, Sol, Terra, and Luna, with Sol positioned as the flagship option for complex reasoning and coding work, Terra as a balanced everyday tier at roughly half the cost, and Luna as a budget option built for speed. Codex now recommends GPT-5.6 Sol as its default model, and OpenAI has confirmed that GPT-5.4 and GPT-5.4 mini will be retired from Codex entirely for ChatGPT-authenticated users by the end of August, with GPT-5.6 Terra and Luna serving as their direct replacements. Independent benchmark tracking has shown GPT-5.6 Sol scoring around 89.5 percent on Terminal-Bench 2.1 at its highest reasoning effort setting, putting it within half a point of Anthropic’s competing Claude Opus 5 on the same test, a sign of just how tightly matched the top coding agents have become this year.
That competitive pressure is likely part of why OpenAI has been shipping features at this pace. Anthropic, Google, and a growing field of coding-focused startups have all been pushing their own agentic development tools aggressively, and benchmarks like SWE-bench Verified and Terminal-Bench have become a kind of running scoreboard the major labs are visibly racing against. Beyond raw model capability, OpenAI has also been investing in the plumbing around Codex itself, recent updates to the open-source Codex CLI have added configurable grace periods for discovering tools from optional MCP servers, allowed extensions to inspect or modify MCP tool results before they reach the model, and tightened sandbox enforcement and permission handling across the app, changes that matter more to engineering teams running Codex at scale than to a casual user asking it to fix a single bug.
OpenAI has also been part of a broader industry effort to standardize how AI agents interact with tools and each other. Earlier this month, Google’s A2A protocol formally joined the Agentic AI Foundation under Linux Foundation governance, aligning it alongside Anthropic’s Model Context Protocol within the same neutral standards body, which now counts more than 250 members including AWS, Microsoft, Bloomberg, and OpenAI itself. A separate agent-plugin packaging standard, agreed on jointly by OpenAI, Amazon, Cursor, Microsoft, and Vercel, has already shipped support across ChatGPT, Codex, GitHub Copilot, Cursor, and VS Code, a notable example of direct competitors agreeing on shared infrastructure rather than each building a walled garden.
Taken as a whole, this stretch of announcements says less about any single flashy feature and more about where OpenAI thinks developer tooling needs to go next. The Admin plugin targets the operational overhead of running Codex and ChatGPT Work at organizational scale. WebMCP and Site tools aim to make agents useful against arbitrary websites rather than just OpenAI’s own products. And the steady model upgrades keep Codex competitive against a field of rivals that are moving just as fast. For engineering teams evaluating which AI coding assistant to standardize on, the practical question isn’t really which model scores highest on a given benchmark this month, it’s which platform is building the surrounding infrastructure, permissions, integrations, and standards support, that will make the tool genuinely usable across an entire organization rather than just for individual developers experimenting on their own. Full documentation on these updates is available through OpenAI’s developer platform, and details on the Admin plugin can be found on OpenAI’s official announcement page.
