ChatGPT Work vs Codex vs Claude Code: What Changed and Which One Should You Use?
OpenAI just turned ChatGPT into a broader work agent. Here is how ChatGPT Work, Codex, Claude Code, Gemini CLI, and Antigravity now fit together.
Last updated: July 10, 2026
Quick answer
ChatGPT Work is the fresh search opportunity. OpenAI launched it on July 9, 2026 as the new way ChatGPT moves from answering questions to doing longer work across web, mobile, desktop, and connected tools. The important detail: OpenAI says Codex technology is built into this broader ChatGPT Work direction, so the old question, "ChatGPT vs Codex," is now too narrow.
Use ChatGPT Work when the job is cross-app business execution: research, files, spreadsheets, meetings, browser tasks, and turning scattered context into a finished deliverable.
Use Codex when the work is technical execution: repo-aware changes, CLI or IDE work, GitHub issues, pull requests, tests, and background engineering tasks.
Use Claude Code when the repository is the operating system: terminal control, local files, project memory, CLAUDE.md, MCP servers, hooks, skills, and rigorous long-running agent workflows.
Use Gemini CLI when your team already lives in Google developer tools. Use Antigravity when you want Google's agent-first IDE/browser/terminal direction rather than a classic coding assistant.
If you are a founder or operator, the best default stack is simple: ChatGPT Work for business tasks, Codex for OpenAI-native technical tasks, and Claude Code for deep repo execution. If you want the operating system behind that stack, start with the AI Chief of Staff resource or the Second Brain AI system.
| Tool | Best use | Do not use it as |
|---|---|---|
| ChatGPT Work | Cross-app business work, research, files, spreadsheets, documents, and delegated desktop/web tasks | A precise repo-native engineering system |
| Codex | OpenAI-native coding, CLI/IDE work, cloud tasks, GitHub issue-to-PR workflows, and technical execution | A general operator workspace for every non-code task |
| Claude Code | Deep local repo work, terminal commands, tests, hooks, memory files, MCP, and repeatable agent workflows | A low-friction tool for non-technical teams with no repo structure |
| Gemini CLI | Google-centered terminal and developer workflows, especially around Google Cloud | The primary choice if your stack is not Google-heavy |
| Antigravity | Google's agent-first command center across IDE, browser, terminal, SDK, and managed agents | A stable, boring replacement for mature repo workflows |
The ranking logic for today: lead with ChatGPT Work because the market is still naming the category. Capture the new query before search volume appears in keyword tools, then bridge it into already-visible demand around chatgpt codex, codex vs claude code, and codex cli.
First, fix the naming problem
People compare these tools as if they were five versions of the same app. They are not.
| Name | What it really is | Primary surface |
|---|---|---|
| ChatGPT Work | OpenAI's broader work surface for longer tasks across chat, files, apps, browser, desktop, and agentic workflows | ChatGPT across web, mobile, desktop, and connected work tools |
| Codex | OpenAI's technical execution layer for code and delegated software work | ChatGPT plan access, Codex cloud, CLI, IDE extension, GitHub |
| Claude Code | Anthropic's repo-native agent for terminal, local files, IDE, web, and automation | Terminal, IDE, Claude Desktop Code, web/cloud sessions |
| Gemini CLI | Google's terminal developer agent for Gemini-centered workflows | Terminal, Google developer tools, Google Cloud context |
| Antigravity | Google's agentic development platform and managed agent stack | Antigravity IDE, CLI, SDK, browser, Gemini API managed agent |
This page compares work surfaces, not raw model intelligence. The best tool is the one whose permissions, context, runtime, and review loop match the work.
People compare these tools as if they were six versions of the same app. They are not.
| Name | What it really is | Primary surface |
|---|---|---|
| ChatGPT | General AI workspace with chat, files, apps, agent mode, deep research, and workspace agents | Browser, mobile, desktop, Slack for workspace agents |
| Codex | OpenAI coding agent that reads, edits, runs code, and can create pull requests | Cloud, CLI, IDE extension, GitHub |
| Claude Code | Anthropic codebase agent for terminal, IDE, desktop Code tab, web, and automation | Terminal, IDE, Claude Desktop Code, web |
| Claude Cowork | Anthropic desktop workspace for longer agentic work, local files, projects, Chrome, and Dispatch | Claude Desktop Cowork tab |
| Gemini | Usually Gemini Code Assist or Gemini CLI when developers say it in this context | IDE, Cloud Shell, terminal, Google Cloud, GitHub |
| Antigravity | Google's agentic development platform and managed agent stack | Antigravity 2.0, IDE, CLI, SDK, Gemini API managed agent |
This page compares work surfaces, not raw model intelligence. The best tool is the one whose permissions, context, and runtime match the work.
The July 2026 state of the market
The category shifted on July 9, 2026. ChatGPT Work makes the old "chatbot vs coding agent" framing obsolete because OpenAI is now packaging Codex-style execution inside a broader work surface.
That creates three layers:
First, general AI workspaces are becoming execution surfaces. ChatGPT Work is the clearest example: the pitch is no longer just "ask ChatGPT" but "delegate ambitious work across context, files, and tools." This is the page's freshness advantage.
Second, technical agents are becoming execution backends. Codex still matters because people search for it directly, developers use it for repo work, and OpenAI now includes Codex across ChatGPT plans. The practical question is whether to send the technical slice of the work to Codex, Claude Code, Gemini CLI, or Antigravity.
Third, repo-native systems still win when correctness matters. ChatGPT Work may be the front door, but Claude Code remains the better fit when the task depends on branch discipline, local tests, memory files, hooks, MCP servers, and repeatable engineering workflows.
So the answer is not "one tool replaces the others." The answer is orchestration: ChatGPT Work for broad delegation, Codex for OpenAI-native technical execution, Claude Code for deep repo work, Gemini CLI for Google-centered developers, and Antigravity for Google's agent-first platform.
Feature matrix
| Capability | ChatGPT | Codex | Claude Code | Claude Cowork | Gemini Code Assist / CLI | Antigravity |
|---|---|---|---|---|---|---|
| Main job | General knowledge work | Software engineering tasks | Repo-native engineering | Longer local desktop work | IDE and Google Cloud coding help | Agent orchestration and managed agents |
| Best first prompt | "Research this, compare options, draft the plan" | "Fix this issue and open a PR" | "Read this repo, make the change, run tests" | "Organize this project and produce the deliverable" | "Help me implement this in my IDE or Cloud Shell" | "Launch an agent workflow across code, terminal, browser, or sandbox" |
| Codebase awareness | Good through GitHub, uploads, files, and app connections | Native repository access through Codex cloud, CLI, IDE, GitHub | Native through local repo, `CLAUDE.md`, memory, MCP, tools | Good for attached local folders and Cowork projects | Good in IDE, local workspace, Cloud Shell, GitHub review | Strong in Antigravity IDE, CLI, SDK, and managed sandbox |
| Writes code | Yes, but less repo-operational by default | Yes | Yes | Yes when routed to Code or given local folders | Yes | Yes |
| Runs commands | Limited by selected agent/tool mode | Yes in CLI, IDE, and cloud environments | Yes in CLI, Desktop local sessions, SSH, and cloud sessions | Yes inside approved local/sandbox workflows | Yes through Gemini CLI and agent mode tools | Yes through IDE, CLI, SDK, and managed Linux sandbox |
| Runs in cloud | Yes for agent and workspace-agent flows | Yes, Codex cloud | Yes, web and cloud sessions | Mostly local desktop, with Dispatch and routed sessions | Cloud Shell and Google Cloud surfaces | Yes, including managed Gemini API sandbox |
| Runs locally | Desktop app, but not a repo-native shell agent | CLI and IDE extension | CLI, IDE, Desktop local sessions | Desktop local folders and projects | Gemini CLI, IDE extensions | IDE, CLI, SDK |
| Background work | Agent and workspace agents can work asynchronously | Cloud tasks run in the background and in parallel | Background agents, routines, web/cloud sessions, Desktop sessions | Dispatch splits and runs child tasks | Agent mode and CLI loops, depending on surface | Designed for multi-step autonomous workflows |
| Pull requests | Possible through connected tools, not the main shape | Native path: GitHub issues, pull requests, cloud diffs | Strong through local git, GitHub, Claude Code web, Slack routes | Can route coding work into Code sessions | Gemini Code Assist for GitHub supports review workflows | Can modify code and produce artifacts, exact PR flow depends on surface |
| Non-code work | Strongest overall | Weak fit | Good if the work lives in repo files | Strong | Moderate, mostly developer-context work | Strong for agent workflows, less mature as a daily writing workspace |
| File and document work | Strong | Repo files | Repo and local files | Strong local file posture | IDE/project files | Sandbox files, project files, agent artifacts |
| Browser work | Strong in agent mode | Not the main path | Preview/browser verification in Desktop and web-adjacent workflows | Claude in Chrome can pair with Cowork | Web search/fetch in CLI and Google tools | Browser control is a core Antigravity theme |
| MCP / connectors | Apps and connectors, plus workspace integrations | Provider and environment setup varies by Codex surface | Strong MCP, connectors, skills, plugins, hooks | Integrations, plugins, skills, MCP through Claude ecosystem | MCP servers in Gemini CLI and Code Assist agent mode | Remote MCP servers and custom tools |
| Best user | Founder, operator, analyst, manager, marketer, researcher | Developer or team that wants OpenAI-native engineering delegation | Developer, technical founder, automation-heavy operator | Operator who wants local desktop agents without terminal work | Developer already using Google Cloud or Gemini Code Assist | Builder who wants Google's agent-first stack |
| Main weakness | Less precise for repo-native execution than dedicated coding agents | Less useful for broad non-code workflows | Requires comfort with files, permissions, and development concepts | Desktop dependency and plan availability matter | Individual-tier migration pressure after June 18, 2026 | Fast-moving preview/platform surface, more to learn |
Which one should you choose?
Choose ChatGPT when the work is bigger than code
ChatGPT is the best default if your task combines research, writing, files, online actions, and business context. It is also the easiest recommendation for non-technical teams because the interface maps to normal work: ask, upload, connect, research, draft, revise, act.
Use ChatGPT for:
- market research and synthesis
- vendor comparison
- slide and document drafting
- spreadsheet edits
- email and workflow tasks
- connected-app reasoning
- general task delegation through agent mode
- repeatable Business or Enterprise workspace agents
Do not force ChatGPT to be your main repo worker if the work needs repeatable tests, branch discipline, terminal commands, and pull requests. That is where Codex or Claude Code fit better.
Choose Codex when you want OpenAI-native engineering delegation
Codex is OpenAI's coding agent for actual software work. The official Codex docs describe cloud tasks that can run in the background and in parallel, repository access through GitHub, IDE delegation, pull request workflows, CLI usage, and configurable cloud environments.
Use Codex for:
- GitHub issue to pull request workflows
- parallel implementation tasks in cloud environments
- OpenAI model access from CLI and IDE
- code review and bug fixing
- applying diffs locally after cloud work
- teams already standardized on ChatGPT workspace permissions
Codex is especially strong when you want a cloud teammate: give it a task, let it work in its own environment, review the diff, and decide whether to merge.
Choose Claude Code when the repository is the operating system
Claude Code is strongest when your work is encoded in files. It reads the codebase, edits files, runs commands, follows CLAUDE.md, uses MCP servers, loads skills, applies hooks, and can move across terminal, IDE, Desktop, web, Slack, and scheduled routines.
Use Claude Code for:
- large repo modifications
- test-first implementation
- local scripts and shell automation
- MCP-heavy workflows
- personal or company coding conventions stored in files
- agent teams, subagents, background tasks, routines, and repeatable workflows
- developer systems where the file tree is the source of truth
This is the tool I would pick for a serious Second Brain-style repository, because the whole system benefits from persistent context files, skills, agents, and terminal access.
Choose Claude Cowork when the task belongs on your desktop
Claude Cowork uses the same agentic architecture as Claude Code, but it is aimed at broader desktop work. It can read and write local files, work inside projects, coordinate sub-agents, pair with Claude in Chrome, and use Dispatch for background tasks.
Use Cowork for:
- research packs
- organizing local folders
- document production
- recurring project work
- tasks that combine files and websites
- delegating from mobile into a desktop session through Dispatch
The key difference from Claude Code: Cowork is less about "make this repo pass tests" and more about "complete this outcome across my local workspace."
Choose Gemini Code Assist or Gemini CLI when you live in Google developer tools
Gemini Code Assist still matters for teams on Google Cloud and IDE workflows. Gemini CLI is an open source terminal agent with a reason-and-act loop, built-in tools, and MCP support. Gemini Code Assist agent mode brings a subset of CLI functionality into VS Code and related IDE surfaces.
Use Gemini when:
- your code already lives close to Google Cloud
- you want Gemini Code Assist in VS Code or JetBrains
- you need Google Cloud-aware coding help
- you use Cloud Shell
- your enterprise already buys Gemini Code Assist Standard or Enterprise
For individuals, treat this as a transition zone. Google says Gemini Code Assist IDE Extensions and Gemini CLI stop serving requests for individual, Google AI Pro, and Google AI Ultra tiers starting June 18, 2026. Antigravity is the migration path.
Choose Antigravity when you want Google's agent-first platform
Antigravity is Google's agentic development platform, not just another code autocomplete product. The docs describe Antigravity 2.0 as a command center, plus IDE, CLI, SDK, and a managed Antigravity agent on the Gemini API. The managed agent can reason, execute code, manage files, browse the web, use Google Search and URL context, register remote MCP servers, and run inside a secure Linux sandbox hosted by Google.
Use Antigravity for:
- multi-step agent workflows
- browser plus terminal plus editor tasks
- managed sandbox execution
- Gemini API agent experiments
- teams evaluating Google's future development stack
- migration away from individual Gemini Code Assist or Gemini CLI tiers
The caution: Antigravity is moving quickly. Some surfaces are preview, and Google documents limitations around structured output, unsupported tools, and remote MCP transport. It is exciting, but you need more tolerance for product motion.
My recommendation
If you are a solo operator or founder, use ChatGPT for broad business work and Claude Code for your repository. Add Claude Cowork if you want desktop delegation and Dispatch.
If your team already pays for OpenAI Business or Enterprise, evaluate Codex seriously. The cloud workflow is clean, especially for issues, pull requests, and parallel engineering tasks.
If your team is Google Cloud-heavy, evaluate Antigravity now. Keep Gemini Code Assist where enterprise IDE support matters, but do not ignore Google's Antigravity migration signal.
If you want one default for building a personal AI operating system, choose Claude Code. It has the best match for file-based memory, local tools, agent skills, hooks, MCP, and repeatable workflows.
If you want one default for company-wide knowledge work, choose ChatGPT. It is easier to roll out across functions, less terminal-shaped, and better for documents, analysis, and everyday business tasks.
FAQ
Is Codex now part of ChatGPT?
Yes, in practical buying and usage terms. OpenAI says Codex is included across ChatGPT plans and that ChatGPT Work has Codex technology built in. Codex is still a distinct coding agent and client surface, but the user journey is increasingly tied to ChatGPT accounts, plans, workspace permissions, and agentic usage.
What changed with ChatGPT Work?
ChatGPT Work changed the frame from chatbot to work agent. The fresh query is not just "ChatGPT vs Codex" anymore. It is "where does ChatGPT Work end, where does Codex begin, and when should I still use Claude Code?" The short answer: ChatGPT Work is the business-work surface, Codex is the OpenAI technical execution layer, and Claude Code is the deeper repo-native system.
Does ChatGPT Work share usage limits with Codex?
For Plus and Pro users, OpenAI says supported agentic features including Codex, ChatGPT Work, and ChatGPT for Excel can draw from the same agentic usage allowance. Codex limits still vary by plan, workspace role, and credits, so check the Codex usage panel or your ChatGPT plan settings before planning a heavy work session.
Should I use Codex or Claude Code?
Use Codex when you want OpenAI-native coding delegation through ChatGPT, Codex cloud, CLI or IDE workflows, GitHub issues, and pull requests. Use Claude Code when your codebase is the operating system and you need local terminal control, tests, hooks, memory files, MCP servers, and repeatable agent workflows. Many technical founders should use both: Codex for parallel OpenAI tasks, Claude Code for the main repo loop.
Can non-developers use Codex?
Yes, but only when the work has a technical shape. OpenAI has been positioning Codex for knowledge work beyond software, including reports, spreadsheets, presentations, contracts, automation, and lightweight tools. Non-developers should start from ChatGPT Work, then use Codex when the task needs technical execution rather than just writing, research, or planning.
What is ChatGPT Work best for?
ChatGPT Work is best for cross-app business execution: research, summaries, documents, spreadsheets, browser tasks, meeting follow-up, workflow automation, and turning messy context into a finished deliverable. It is the better starting point for operators, founders, analysts, and managers.
What is Claude Code still best for?
Claude Code is still best when the repo itself holds the operating system: source files, scripts, tests, project instructions, memory files, MCP servers, skills, hooks, and deployment workflow. If correctness depends on reading the codebase, making edits, running commands, and reviewing diffs, Claude Code remains the stronger default.
Where do Gemini CLI and Antigravity fit?
Gemini CLI fits Google-centered developers who want terminal workflows around Gemini and Google Cloud. Antigravity is Google's agent-first platform direction across IDE, terminal, browser, SDK, and managed agents. They are worth evaluating if your stack is Google-heavy, but they are not the first answer for a ChatGPT Work searcher unless Google infrastructure is already central.
Which one should a non-technical founder start with?
Start with ChatGPT Work. Add the AI Chief of Staff workflow if you want repeatable delegation. Add Second Brain AI if you want the memory and operating system around those agents. Bring in Codex or Claude Code only when your business work becomes technical execution.
Which one should a technical founder start with?
Start with Claude Code or Codex for the repo, then use ChatGPT Work for research, GTM, docs, and broad synthesis. If you want one practical setup: Claude Code runs the main repository loop, Codex handles extra OpenAI-native engineering tasks, and ChatGPT Work turns the output into business deliverables.
Official sources checked
- OpenAI: ChatGPT is now a partner for your most ambitious work
- OpenAI Help: Using Codex with your ChatGPT plan
- OpenAI Help: Using Credits for Flexible Usage in ChatGPT
- OpenAI: Codex is becoming a productivity tool for everyone
- OpenAI: Introducing workspace agents in ChatGPT
- OpenAI Developers: Codex cloud
- OpenAI Developers: Codex workflows
- OpenAI Developers: Codex IDE extension
- Claude Code overview
- Gemini CLI
- Google Antigravity overview
Bottom line
Use ChatGPT Work to run work across the business.
Use Codex to delegate technical execution inside OpenAI's ecosystem.
Use Claude Code to operate a codebase like a persistent working system.
Use Gemini CLI when Google developer infrastructure is already the path.
Use Antigravity when you want Google's agent-first future and can tolerate a faster-moving platform.
For most founders, the compounding move is not choosing one tool. It is building the operating system that routes the right job to the right agent. Start with the AI Chief of Staff resource if you want the workflow, or Second Brain AI if you want the full memory system.