OpenAI Workspace Agents, Codex, and Images 2.0 Explained

Remember when ChatGPT was a chat box you typed into and got text back? That era is officially ending, and OpenAI just spent seven days driving the point home. In a single week, they shipped a Codex desktop app that controls your Mac with its own cursor, an image model that searches the web and thinks before drawing, and Workspace Agents: shared, Codex-powered AI coworkers teams can build once and run across ChatGPT and Slack.

Individually, any one of these is a big release. Together, they signal exactly what OpenAI thinks the next phase of AI looks like: not smarter models answering prompts, but agents running your team's actual workflows.

First up, the TL;DR

ChatGPT is quietly morphing into an operating system for work. Here's the whole launch wave at a glance.

Here's what happened:

Today, April 22: Workspace agents launched in research preview. They're Codex-powered agents (meaning they run on OpenAI's cloud, so they keep working when you log off) that teams can build once and share across ChatGPT and Slack. Available on ChatGPT Business, Enterprise, Edu, and Teachers plans. Free until May 6, 2026, then credit-based pricing.
April 21: ChatGPT Images 2.0 (model name: gpt-image-2) rolled out to all ChatGPT users, including free. It's OpenAI's first image model with built-in reasoning, meaning it can search the web, plan, and self-check before generating. Native support for Japanese, Korean, Hindi, Bengali, and more.
April 16: Codex for (almost) everything added "background computer use" (Codex operates your Mac with its own cursor, in parallel with your own work), an in-app browser, image generation via gpt-image-1.5, persistent memory, 90+ new plugins, and "Heartbeat Automations" that wake up across days or weeks.

How to try it:

Workspace Agents: In ChatGPT Business/Enterprise/Edu, click Agents in the sidebar, then Create. Describe the workflow in plain English; the builder wires up steps, tools, and triggers. Connect a Slack channel from agent settings if you want it to respond there too.
Images 2.0: Open any ChatGPT chat and prompt for an image. Plus, Pro, Business, and Enterprise users can toggle Thinking mode for web search, 8-image consistency, and self-checking.
Codex computer use: Download the Codex desktop app (full cursor control on Mac, rest on Windows). On macOS, enable Codex under System Settings > Privacy & Security > Accessibility and Screen Recording, then flip on Computer Use in app settings.

Why this matters: The frontier just moved from "models that answer" to "agents that do." OpenAI is racing Anthropic (which shipped Claude Code Routines eight days ago) and Google (which launched Deep Research Max the same week) for ownership of the entire work layer, not just the chat window. For your org, the big Q3 question isn't "which chatbot do we buy?" It's "whose agents own our workflows?"

Our take: Bundle all three launches and the pattern is obvious: OpenAI wants to be the place your team's workflows live, not just the place people ask questions. The open question is governance. Will enterprise IT trust an agent with memory, cursor-level access, and 90+ connected apps? Role-based controls and the Compliance API help, but the blast radius of one compromised or misconfigured agent has never been larger.

How we got here: the agent shift in 2026

Up until this year, AI was something you had conversations with. You'd ask a question, get an answer, do the work yourself. Useful, but limited.

Three things changed through late 2025 into 2026:

Models got good enough to plan multi-step work. GPT-5.3, Claude Opus 4.7, and Gemini 3 Pro crossed the threshold where you could hand them a goal instead of a prompt.
Computer use matured. Anthropic pioneered it in late 2024; Claude shipped native Mac control in March 2026; OpenAI matched it in Codex on April 16.
Enterprises started demanding shared workflows. "One user, one prompt" was fine for experimentation, but teams needed agents that followed company processes, used approved tools, and produced consistent outputs.

Workspace Agents, Routines, and Deep Research Max are all answers to the same question: how does AI graduate from individual productivity tool to team infrastructure?

What is a workspace agent, exactly?

OpenAI's Academy guide defines an agent as a system that carries out a task with three parts: a trigger (what starts it), a process (the steps it follows, sometimes using specialized "skills"), and tools (the approved systems it can read from or write to).

Workspace agents specifically are Codex-powered and cloud-hosted. The three pillars:

Build once, share across the org. Agents appear in a Team directory in the ChatGPT sidebar, alongside agents you've built yourself and ones you've recently used.
Run on triggers. Agents can run on a schedule (say, every Friday at 9am), be @-mentioned in a ChatGPT conversation, or listen for messages in a connected Slack channel.
Govern by role. Admins control who can build, publish, and run agents; which connectors each can access; and which write actions (sending email, posting in Slack, updating records) require human approval. Every agent's configuration, updates, and runs are visible through the Compliance API.

OpenAI shipped five templates to start with:

Software Reviewer: triages app requests against IT policy, files tickets when needed.
Product Feedback Router: turns Slack and support feedback into prioritized tickets and weekly summaries.
Weekly Metrics Reporter: pulls data every Friday, generates charts, drafts narrative, delivers the report.
Lead Outreach Agent: qualifies inbound leads against your rubric, drafts personalized follow-up emails, updates the CRM.
Third-Party Risk Manager: screens vendors for sanctions, financial, and reputational risk, produces a structured report.

The quote from Rippling in the press release is the one to pay attention to: a Sales Consultant (not an engineer) built a Sales Opportunity agent end to end. It now researches accounts, summarizes Gong calls, and posts deal briefs to Slack, saving reps 5-6 hours a week on every deal. That's the real pitch: agents your non-technical staff can ship.

What's actually new in Codex for (almost) everything

OpenAI's April 16 update turned Codex from a developer-only coding agent into something closer to a full desktop automation layer, serving more than 3 million weekly developer users.

The headline feature is background computer use: on macOS, Codex operates other applications with its own cursor, clicking and typing across Chrome, Slack, Figma, your design tools, whatever, while you keep working in parallel in other apps. Windows users get computer use too, but without the cursor-level background interaction at launch. EU, UK, and Switzerland users are excluded from computer use for now.

Four other updates worth knowing about:

90+ new plugins bundle skills, app integrations, and MCP servers (Model Context Protocol, a vendor-neutral standard for connecting agents to tools). Notable names: Atlassian Rovo, GitLab Issues, Microsoft Suite, Neon by Databricks, CircleCI, CodeRabbit, Remotion, Render.
In-app browser lets you comment directly on live web pages to give the agent precise instructions. Useful for front-end iteration today; likely to expand beyond localhost over time.
Memory (preview) stores your preferences, corrections, and project knowledge across conversations, so Codex learns your tech stack and style. Not yet available in EU, UK, Enterprise, or Education accounts.
Heartbeat Automations let Codex schedule future work for itself, "wake up" to resume a task days or even weeks later, and proactively propose work based on your memory and connected plugins. This is OpenAI's answer to Anthropic's Routines.

One thing worth flagging: in December 2025, BeyondTrust researchers found a critical command injection flaw in Codex's CLI where a malicious GitHub branch name could leak a user's auth token. OpenAI patched it in February 2026, but it's a reminder that "agent that operates your computer" means the blast radius of one bug is larger than ever. Sensible best practice per AI Automation Global: run Codex under a dedicated macOS user account isolated from your actual work account, especially for sensitive workflows.

Why Images 2.0 is different

ChatGPT Images 2.0, which OpenAI announced April 21, is the first image model in the OpenAI lineup with built-in reasoning. In plain English: before it generates anything, it can think about what you asked, search the web for current information (logos, recent events, cultural references), draft multiple attempts, and self-check the result.

The practical wins for knowledge workers:

Dense text actually renders. Full paragraphs in Japanese, Korean, Hindi, Bengali, Chinese, Arabic. Historically the single hardest thing for image models; now close to production-ready. PetaPixel's early review called this out as the headline improvement.
8-image consistency. Generate a set of 8 images from one prompt with characters and objects preserved across them. VentureBeat called this "character and object continuity" across the series. This is what kills the "I've generated 12 variations and the character's face keeps changing" problem for storyboards and social graphics.
Thinking mode with web search (Plus, Pro, Business, Enterprise only). The model can look up current information before generating, so it won't invent a 2024-era logo for a company that rebranded last quarter.
Pricing tiers for developers. Roughly $0.006 for low quality, $0.053 for medium, $0.211 for high at 1024×1024 via the API (as gpt-image-2), per Digital Applied's breakdown.

What's notable here isn't any one feature; it's that an image model is now functionally an agent. It plans, researches, generates, and verifies. The same paradigm OpenAI is rolling out for text and code workflows is now standard for pixels too.

The competitive frame: OpenAI vs Anthropic vs Google

Look at the sequence of events:

April 14, 2026: Anthropic launches Routines in Claude Code (cloud-hosted, scheduled, API- or GitHub-triggered) plus a redesigned Claude Code desktop app with parallel agent sessions.
April 16: OpenAI ships Codex for (almost) everything.
April 20-21: Google ships Deep Research Max, its most powerful autonomous research agent.
April 21: OpenAI ships Images 2.0.
April 22 (today): OpenAI launches Workspace Agents.

Each company is racing to own a different layer of the agent stack.

Anthropic's bet is that individual power users (developers, technical knowledge workers) adopt Claude Code and Routines first, then enterprises follow organically. OpenAI's bet is the opposite: win enterprises directly with shared, governed team agents, and let individual ChatGPT usage ride along. Google is betting that depth (Deep Research Max's multi-hour autonomous research runs) matters more than breadth.

For buyers, this is great news. Whichever vendor you prefer, you're about to get a meaningfully better agent product because the competitive pressure is absurd. The concern is vendor lock-in: OpenAI's plugin format is proprietary, while MCP is vendor-neutral. Teams that want to stay portable should lean on MCP servers for any custom integrations rather than the plugin catalog.

What this means for you

Three concrete moves based on where you sit:

If you're on ChatGPT Business, Enterprise, Edu, or Teachers: Try the Workspace Agents research preview before May 6, while it's free. Start with one of the templates (Weekly Metrics Reporter is the highest-leverage first build for most teams, per OpenAI's own examples). Put the agent in a low-stakes Slack channel first; watch how it handles edge cases before graduating it to anything important.
If you're technical or technical-adjacent: Install the new Codex desktop app, turn on computer use, and give it one repeatable task you hate doing (research synthesis across tabs, triaging a Jira backlog, writing status updates from Slack threads). The goal is learning what it can and can't do on your machine before you design real workflows around it.
If you produce visual content: Test ChatGPT Images 2.0's Thinking mode against whatever you currently use (DALL-E, Midjourney, Google's Nano Banana 2, Flux). Pay special attention to text rendering if you produce slides, social graphics, or infographics. The text-accuracy delta is real.

And the open question every team should be wrestling with by Q3: who on your team owns agent governance? If the answer is "nobody yet," you already have a problem; the agents are arriving whether or not you have a policy for them.

Where this goes next

OpenAI's roadmap hints at what's coming: admins seeing every agent built across their org in the admin console, new trigger types beyond schedule and @-mention, workspace agents in the Codex app, better performance dashboards. Translation: everything announced this week is the rough draft. The real question is whether your team figures out how to use it before your competitors do.