What each product actually is
Jules. Google's documentation describes Jules as a coding agent that helps you fix bugs, add documentation, update your app and implement new features. It integrates with GitHub and works autonomously, so you can submit a task, do something else and return when it is done. You sign in with a Google account, connect GitHub, choose a repository and branch, and write a prompt. You can also assign work by adding the "jules" label to a GitHub issue. Jules clones the repository into a cloud VM, proposes a plan, shows a diff and opens a pull request.
Codex. OpenAI's documentation presents Codex as a set of surfaces. The Codex CLI lets you inspect code, make changes, run commands and automate repeatable work without leaving your terminal, against your local repository and the tools already installed on your machine. Codex Cloud runs coding tasks remotely: you create or select an environment, Codex prepares your repositories, and each task gets its own workspace that can keep working while your computer is asleep. The CLI source is published on GitHub at openai/codex under the Apache-2.0 license.
Current surfaces (per vendor pages)
- Jules: web app at jules.google, GitHub integration including issue labels, Jules Tools CLI (changelog entry: Oct 2, 2025), Jules REST API (Oct 3, 2025), scheduled tasks and suggested tasks, and MCP support (Feb 2, 2026).
- Codex: CLI for macOS, Linux and Windows (standalone installer, npm or Homebrew), IDE extension, desktop app, web and iOS (listed on the Plus plan), Codex Cloud, and codex exec for scripts and CI.
- Both: GitHub-connected cloud tasks that end in reviewable changes and pull requests.
Scoreboard (from existing aicoolies reviews)
The aicoolies reviews score Jules at 83 overall (speed 85, privacy 68, developer experience 85; tested 2026-05-05) and Codex at 80 overall (speed 72, privacy 68, developer experience 79; tested 2026-03-25). Both products have shipped significant changes since those tests, so treat the scoreboard as dated secondary evidence. It does not settle the comparison on its own.
Workflow: async cloud VM vs local-first plus cloud
The deepest difference is where the agent runs by default. Jules always works in the cloud. Each task runs in a fresh, short-lived virtual machine where Jules clones your repository, installs dependencies and makes changes based on your prompt. For simple projects Jules studies the repository and refers to AGENTS.md or the README to set up the environment. For complex ones you provide a setup script. Long-running processes such as dev servers or watch scripts are not supported in setup scripts, so you use discrete install and test commands. Environment snapshots (changelog: Aug 5, 2025) speed up repeated tasks.
Codex starts on your machine. Run codex in a project directory, sign in with ChatGPT or another method, and the agent can explore code, plan a change, edit files and run your local tools in the same session. You can steer the active turn and inspect commands and diffs as they appear. When work should run remotely, Codex Cloud uses published environments: Codex inspects your repositories, installs dependencies, tests the workflow and asks for missing access. You review and publish the environment, then start tasks from it. From the terminal, codex cloud lets you browse cloud chats, submit work to a configured environment and apply the result to your local repository.
In practice, Jules suits a queue of delegated GitHub tasks that you review later, while Codex suits developers who alternate between hands-on terminal work and delegated cloud work on the same codebase.
Planning, review and control
Jules puts the plan front and center. It proposes a plan you can approve or adjust, and the changelog lists an interactive plan and a critic agent (Aug 8, 2025), a planning critic for auto-approved plans (Jan 26, 2026), the ability to act on PR feedback (Sep 23, 2025) and automatic fixing of CI failures (Feb 19, 2026). You can leave while Jules works and get browser notifications when a plan is ready or a task completes.
Codex emphasizes explicit permissions. In the CLI, the /permissions command lets you choose when Codex can edit files or run commands without asking, and lets you inspect the active sandbox and writable roots before continuing. OpenAI also recommends creating Git checkpoints before and after a task so you can revert changes. In Codex Cloud, you review changed files and test results, request follow-ups, and commit or open a pull request when ready.
Licensing and billing
Jules. Jules has a no-cost plan plus paid tiers delivered through Google AI plans. Google's limits page lists 15 daily tasks and 3 concurrent tasks on the base Jules plan, 100 daily and 15 concurrent with Jules in Pro (included in Google AI Pro), and 300 daily and 60 concurrent with Jules in Ultra (included in Google AI Ultra). Limits use a rolling 24-hour window. Paid Jules plans currently require an individual Google account ending in @gmail.com; business users are pointed to an interest form. On Google's Gemini subscriptions page, Google AI Pro is listed at $19.99 per month and Google AI Ultra from $99.99 per month (US prices).
Codex. OpenAI's pricing page lists Free ($0) and Go ($8 per month) with GPT-6 Luna in the desktop app, subject to rollout. Plus ($20 per month) covers Codex on the web, CLI, IDE extension and iOS, along with cloud integrations such as automatic code review and Slack. Pro plans cost $100, $200 or $500 per month. Business is $20 per user per month billed annually for 2+ users, or $25 billed monthly, and Enterprise and Edu are also available. API-key usage covers the CLI, SDK and IDE extension and is billed at API pricing, but it excludes cloud features such as GitHub code review and Slack.
Who should pick which
- Pick Jules if your work already lives in GitHub, you like assigning issues via labels, and predictable daily task counts matter more than token-level billing.
- Pick Jules if you already pay for Google AI Pro or Ultra, because Jules in Pro and Jules in Ultra are included as plan benefits.
- Pick Codex if you want an agent in your terminal with your own toolchain, permission controls and a sandbox you can inspect.
- Pick Codex if you need CI automation with API keys, because codex exec and API-key billing are documented paths.
- Pick Codex if your team buys ChatGPT Business or Enterprise and wants admin controls such as SAML SSO and MFA in a dedicated workspace.
Decision guide: five practical forks
- Where must code execute? Only in a vendor cloud VM points to Jules. Locally first, with optional cloud, points to Codex.
- How do you buy AI? A personal Google AI plan points to Jules. ChatGPT plans or API keys point to Codex.
- Is this a business account? Paid Jules tiers currently target @gmail.com accounts; Codex documents Business and Enterprise plans.
- Do you need CI pipelines? Codex documents codex exec for scripts and CI. Jules offers a REST API and CLI for programmatic use.
- Do you need an open-source client? The Codex CLI repository is Apache-2.0; no open-source agent client is documented on the Jules pages reviewed here.
Feature surfaces that matter in practice
Jules features that matter in daily use include scheduled tasks for routine maintenance and suggested tasks that let Jules find issues proactively (both Dec 10, 2025), memory (Sep 30, 2025), web browsing (Aug 8, 2025), web-app testing with visual results (Aug 7, 2025), environment variables (Oct 1, 2025), stacked diffs and image upload. The VM runs Ubuntu Linux with Node.js, Bun, Python, Go, Java and Rust preinstalled. The FAQ says Jules is language-agnostic but works best with JavaScript/TypeScript, Python, Go, Java and Rust.
Codex features that matter in daily use include choosing the model, reasoning effort, permissions and commands per task, non-interactive codex exec for repeatable workflows, cloud environments with configurable internet access and network secrets, saved state reused across tasks, and Slack integration through ChatGPT. On Plus, OpenAI lists cloud-based integrations such as automatic code review.
Cost and procurement notes
Jules is priced in tasks, Codex in plan usage or tokens. Jules' limits are explicit counts of daily and concurrent tasks per tier. Codex usage depends on the model, task size and complexity, and on whether the task runs locally or in the cloud. OpenAI notes that cloud tasks may use more of your allowance than local messages, and that Pro plans currently have no five-hour limit. Teams that need a purchase order or seat-based admin controls today will find that path documented for Codex Business and Enterprise. For Jules, Google says it is working on upgrade paths beyond individual Google accounts.
Privacy and data-path framing
Jules always runs your code in a cloud VM with internet access, and Google advises treating it like any shared compute surface: do not commit secrets, avoid vulnerable dependencies and be careful with third-party packages. Google states that Jules does not train on private repository content. You control which repositories Jules can see through the Google Labs Jules GitHub app settings.
With Codex you can keep the loop on your own machine through the CLI and decide per run how much the agent may do without asking. Codex Cloud lets you limit an environment's network destinations and use network secrets only for allowed HTTPS services. On Business, OpenAI states there is no training on your business data by default.
Speed and developer experience
Neither vendor publishes a like-for-like latency benchmark between these two products, so this comparison makes no speed claim beyond the dated aicoolies scores above. Experience differs by design. Jules optimizes for parallel, unattended work with up to 60 concurrent tasks on Ultra. Codex optimizes for an interactive loop where you steer turns in real time, with cloud delegation when you need it.
Migration and coexistence patterns
The two tools can coexist. A common split is to use Jules for queued GitHub maintenance such as dependency bumps, test additions and scheduled chores, which it explicitly targets, and to use Codex for interactive feature work and CI automation. Both operate through Git branches and pull requests, so review stays in your existing GitHub workflow either way. Keep AGENTS.md accurate: Jules reads it for environment hints, and it gives any agent a shared source of project conventions.
Practical evaluation checklist
- Run the same three tasks (a bug fix, a test addition and a dependency bump) on both tools against a non-critical repository.
- For Jules, write a setup script and confirm that tests run in the VM without long-running processes.
- For Codex, try both a local CLI session with restricted permissions and a Codex Cloud task from a published environment.
- Compare review effort: plan quality, diff size, PR description and the number of follow-up rounds.
- Map costs: Jules task counts per tier versus Codex plan usage or API tokens for your expected volume.
- Confirm account eligibility: an individual @gmail.com account for paid Jules tiers, or the ChatGPT plan type for Codex features.
Frequently confused points
The Jules plan pages and changelog name different Gemini models at different points in time. The changelog lists Gemini 3 Flash becoming the base model on Jan 30, 2026 and Gemini 3.1 Pro becoming available on Mar 9, 2026, so check the current limits page for the model tied to each tier. Codex model availability likewise depends on plan and surface, and with API keys it follows the API models available to your key.
