Quick answer: which one, and for whom
Two of these are agents you point at a repository. The third is an editor you sit inside. That structural difference decides most of the choice, and it will still be true when every price on this page has changed.
Claude Code is Anthropic's agentic coding tool — in its own words, a tool that "reads your codebase, edits files, runs commands, and integrates with your development tools," available in the terminal, IDE extensions, a desktop app and the browser. It runs Claude models only. Best when you want an agent to take a described outcome and work across many files under permission rules you set first.
OpenAI Codex is a coding agent reachable from a terminal CLI, an IDE extension, Codex cloud and the ChatGPT apps. The Codex CLI is open source under Apache-2.0. It suits companies already standardised on ChatGPT seats, and its sandbox defaults are conservative enough to show a security reviewer.
Cursor is an actual code editor — its migration docs say it "is based upon the VS Code codebase," with one-click import of your VS Code extensions, themes and keybindings. It is the most model-agnostic of the three, and it fits when a human is doing the typing. One ownership change matters: Cursor's blog announced on 14 August 2026 that "Cursor has officially been acquired by SpaceX," so its parent company Anysphere now sits inside the SpaceXAI division.
Pricing and model line-ups change constantly. Every figure below was checked against vendor documentation in September 2026. Check the current pricing page before you budget, and treat any benchmark as perishable.
The structural difference that decides the choice
All three read repositories, edit multiple files, run commands, review diffs and speak the Model Context Protocol. Two architectural differences last.
Terminal coding agent versus AI-native IDE
Claude Code and Codex are agent-first: you describe an outcome, the agent explores the repo, plans, edits files, runs your tests, shows the result. Both grew surfaces around that core — Claude Code has VS Code and JetBrains extensions, a desktop app and a web surface; Codex has a VS Code extension that also runs inside Cursor and Windsurf. Cursor inverts it: the AI lives in the editing surface as inline Tab completion, an Agent panel, Plan Mode and in-place diff review.
The consequence is automation, not taste. Terminal agents are trivially scriptable — pipe a log file into Claude Code, run Codex non-interactively in CI. An editor is better when a human steers minute by minute.
Model-locked versus model-agnostic
Claude Code runs Claude models — the Claude Opus, Claude Sonnet and Haiku tiers. What you can change is where inference happens: Anthropic's API, Amazon Bedrock, Google Cloud's Agent Platform (formerly Vertex AI), Microsoft Foundry, or a self-hosted LLM gateway. That is deployment flexibility, usually what a regulated buyer actually wants. It is not model choice.
Codex is built around OpenAI's GPT-5 series and signs you in with a ChatGPT account. Because the CLI is open source, its config at ~/.codex/config.toml exposes a model_providers table with base_url and env_key fields, so it can be pointed elsewhere — a documented escape hatch, not a multi-vendor product.
Cursor is the multi-vendor option. Its docs list frontier models from OpenAI, Anthropic, Google and SpaceXAI — Claude, GPT and Gemini families in one tool — alongside Cursor's own Composer model. You can add your own keys for OpenAI, Anthropic, Google, Azure or AWS Bedrock, with two caveats from Cursor's help docs: custom keys work for chat models only, so Tab completion still uses Cursor's built-in models, and Cursor's zero-data-retention policy does not apply when you supply your own key. We cover that trade-off in our guide to bring-your-own-key AI and the OpenAI API key.
The acquisition matters in one way: Cursor's model-agnosticism is now a product decision made by a company that also builds a frontier model, and its in-house pool carries the most included usage. The third-party models are still there as of September 2026 — but if cross-vendor choice is why you are buying, check the current model list.
What each one is genuinely best at
Claude Code: unattended work with a memory of your standards
The workflow that suits Claude Code is a described outcome rather than a described edit. "Write tests for the auth module, run them, and fix any failures" is the canonical example from Anthropic's docs: coding tasks spanning several files you must run to know whether they worked.
The feature that matters most to a business is the least glamorous. A CLAUDE.md file in the repository root is read at the start of every session, so coding standards and review checklists travel with the codebase instead of living in one person's head. Hooks run your own shell commands before or after the agent acts, so a formatter or lint gate is enforced rather than hoped for.
Codex: conservative defaults and a bill you already pay
Codex earns its place on an open-source, auditable CLI and a permission model explicit enough to explain to a nervous stakeholder. Its config exposes three sandbox modes — read-only, workspace-write (the default), and danger-full-access — plus an approval_policy of untrusted, on-request or never. In workspace-write, Codex reads, edits and runs commands inside your working directory, and network access is off unless you enable it.
Codex is also included across ChatGPT plans you may already pay for, and the CLI and IDE extension share one ~/.codex/config.toml, so a policy set once applies in both.
Cursor: continuous assistance while a human drives
Cursor wins where work is exploratory and the person wants to feel every step. Tab completion that predicts your next edit is a different category from an agent that disappears for ten minutes. It is also the least disruptive migration, since a developer keeps their VS Code muscle memory — with one note: Cursor pulls extensions from the Open VSX registry rather than Microsoft's marketplace, so a few are unavailable.
Claude Code vs Codex: the head-to-head people search for
Which models do Claude Code and Codex use?
Claude Code uses Anthropic's Claude models and nothing else. Codex uses OpenAI's models and, by default, nothing else. Neither runs the other's: Codex does not use Claude, and Claude Code is not powered by GPT. The confusion is fair, because Cursor sits above both and runs either.
Within each you still pick a tier, and that drives most of your bill. Anthropic's guidance is blunt: Claude Sonnet handles most coding tasks well and costs less than Claude Opus, which should be reserved for complex architecture. Codex exposes the same trade-off through model_reasoning_effort, from minimal through xhigh.
Which has better sandboxing?
Codex makes the sandbox the primary control: one setting, three named modes, network off by default — the more legible model for a business that must document what a tool may do.
Claude Code makes permissions primary and layers a sandbox underneath. In manual mode it starts read-only, runs harmless read commands such as ls, cat and git status without asking, and confines writes to the folder it was started in and its subfolders; curl and wget are not auto-approved. Organisations can push managed settings that local config cannot override.
So Codex has the simpler sandbox story, Claude Code the richer governance story. Neither replaces the real gate, which is git. Worth reading in Codex's security docs is what its network proxy does not filter — web search, connector tool calls, MCP connections, browser activity and cloud tasks among them.
Which is faster, and which produces better code quality?
Do not buy on benchmarks. Both labs ship coding models every few weeks, the leaderboard order flips with each release, and published scores use public test sets that look nothing like your codebase. Speed is mostly a setting anyway: wall-clock time is dominated by the model tier, the reasoning effort you allowed, and how much context the session carries.
The useful test is boring. Take one real ticket from your backlog, run it through both, and compare the diffs rather than the clock. Code quality on your stack with your conventions is the only benchmark that predicts anything, and teams routinely find one tool suits their language better than the other.
Is Claude Code open source? Is Codex?
Codex partly is: the Codex CLI is on GitHub under Apache-2.0, so you can read what it does before running it. Claude Code is not — its repository is a distribution and issue tracker, and its LICENSE.md reads in full: "© Anthropic PBC. All rights reserved. Use is subject to Anthropic's Commercial Terms of Service." If source availability is a hard procurement requirement, that decides it immediately.
Multi-file work, running commands, and code review
Plan first, whichever you pick. Claude Code's plan mode proposes an approach for approval before writing anything; Codex has plan-style scoping and a goal capability for long milestones; Cursor has Plan Mode in the editor. Approving a plan is cheaper than reviewing a wrong implementation, and this habit does more for cost and quality than any model choice.
On code review they converge: read the diff. Claude Code stages changes, writes commit messages, opens pull requests and can run automated review on PRs via GitHub Actions; Codex reviews its own diffs and surfaces them in the editor. In all three, git is the real review gate.
Codex supports MCP servers, and so do Claude Code and Cursor. The Model Context Protocol is how a coding agent reads a Jira ticket or design doc instead of being told about it; if that is new, start with our explainer on connecting AI agents to tools via MCP.
What each costs, and what drives the bill
Cursor is seat-plus-usage. As of September 2026 its pricing docs list Pro at $20/month, Pro Plus at $60/month and Ultra at $200/month, with Teams at $40 and $120 per user per month. Usage splits into a Cursor Models pool — its own Composer and Grok models, with more included usage — and an Other Models pool billed at API rates.
Claude Code is subscription or pay-as-you-go. It is included in the paid Claude plans — as of September 2026 claude.com/pricing lists Claude Pro at $20/month billed monthly, with Claude Max, Team and Enterprise above — or you use an Anthropic Console API key and pay per token, or route through a cloud provider. Codex is included across ChatGPT plans and uses credit-based billing beyond a plan's allowance; check OpenAI's current rate card, because that is the number that moves.
So which is cheaper, Claude Code or Codex? At the entry tier the difference is noise — both are reachable from a roughly $20/month subscription, and if you already pay one vendor, trying the matching agent costs nothing extra. Above that, neither is answerable in the abstract, because both meter usage rather than seats.
What drives spend is context size, not feature count. Anthropic's cost documentation is candid: long sessions never cleared, cache misses, and leaving the most expensive model as the default account for most surprise bills. Its published enterprise figure is around $13 per developer per active day and $150-250 per month, with 90% of users under $30 per active day — an order-of-magnitude anchor, not a quote.
Two habits cut the bill more than any setting: clear the session between unrelated tasks, and match the model to the job rather than running the largest all day. Anthropic also notes that agent teams use roughly seven times the tokens of a standard session.
Security and governance: what a business should ask
What leaves your machine
All three send your code to a model over the network. None works offline. Claude Code sends to Anthropic's API by default, or to your own cloud tenancy, or through a gateway you host; its web sessions run in isolated Anthropic-managed VMs with limited network access, scoped git credentials and audit logging. Codex sends to OpenAI, with telemetry off by default.
Cursor holds SOC 2 Type II, ISO/IEC 27001:2022, ISO/IEC 42001:2023 and AIUC-1, and offers a Privacy Mode under which it does not train on your data. One detail is worth internalising: its help docs state that even when you bring your own API key, requests still route through Cursor's servers for final prompt building. The key is not persisted, but Cursor is in the path. For legal, health or defence work that can be the deciding fact — if it is yours, look at self-hosted AI for business first.
Review gates, repository access, and whose account is billed
Give a coding agent access to one repository, not your whole machine. All three respect a working directory; Claude Code enforces it as a hard boundary in manual mode. Set the review gate at the branch, not the keystroke — approving every individual edit is exhausting, and past the first week nobody does it.
Then answer the billing question explicitly. On a personal ChatGPT or Claude subscription, usage sits on an individual's consumer account under consumer terms. On a Team or Enterprise plan, or a company-owned API key, it sits on the business with commercial terms, central spend caps and per-user reporting. If a contractor does the work, decide before day one whose account it lands on.
Enforce policy centrally rather than trusting each laptop: Claude Code has managed settings local config cannot override plus OpenTelemetry cost export, Codex has admin-restricted hook configuration, and Cursor has enterprise controls including SCIM.
Can you run Claude Code and Codex together on one project?
Yes, and many teams do. They are separate programs operating on the same git repository — each with its own config file, permissions and account — so there is no technical conflict. The only real rule is not to let two agents edit the same files at once, which is the rule you already apply to two humans. Use branches.
The vendors have stopped pretending it is either-or: the Codex extension is explicitly supported inside Cursor, and Anthropic publishes a Claude Code extension install link for Cursor alongside the VS Code one.
The pattern is roles rather than winners — an editor for work a person has their hands on now, a terminal agent for work you describe and walk away from, and often a second agent for code review, because a model that did not write the code is better at finding what is wrong with it. What you should not duplicate is governance. And two things do not work: Claude Code through a ChatGPT subscription, or Codex through a Claude plan.
A decision framework for three real situations
Solo founder with no developer
Pick an agent, not the editor. Cursor's strengths — inline completion, model switching, editor ergonomics — are strengths for someone who can already read code. If you cannot, an editor full of AI features is a cockpit you cannot fly.
Claude Code or Codex, in plan mode, on a small scoped project in a fresh git repository, is a reasonable first experiment. Whether you should run a business on the result is a separate question, and the honest answer is usually no — not because the code is bad, but because you will not know which parts are load-bearing when it breaks at 2am.
A team with one developer
Let the developer choose the tool and standardise the governance — one person's preferences should not become the company's only copy of how the system works. Insist on three things: coding standards live in a file in the repository (CLAUDE.md for Claude Code, the equivalent rules files for Codex and Cursor), agent work lands on branches with pull requests, and the AI account belongs to the business.
An agency-built system you now maintain
Your requirement is comprehension before modification. The first job for any of these tools is not writing code; it is explaining the code you already own — tracing how a request flows, describing what a scheduled job does. That is the highest-value use for a business that inherited a system, and all three do it well.
Ask for a plain-English architecture summary, verify it against reality, and commit it to the repository. Before changing anything, confirm you own the pieces: repository, hosting, domain, database, API keys. No AI tool can fix a system you lack the credentials to reach. That ownership question is the spine of how we structure consulting work — the client holds the accounts, always.
What none of them replace
All three are extremely good at the part of software that was always most visible and least valuable: typing. None is good at knowing what to build.
The failure mode we see most is not bad code. It is a perfectly working system that automates the wrong step — one that was never the bottleneck — while the actual constraint sits untouched upstream. An agent will build that quickly, and it will pass its own tests. Deciding which process is worth automating is still human work. Our guide to what AI agents actually are covers where that line falls.
The second thing none replaces is ownership. Whatever gets built, someone must hold the repository, hosting, keys and the ability to change it next year — which has nothing to do with what wrote the code.
AutomateNexus is an accepted partner in both OpenAI's partner program and Anthropic's Claude Partner Program. We use all three of these tools and have no reason to sell you one over another.
If you want the system rather than the tooling, that is our work: builds start at $7,500, with a typical timeline of about 30 days or four to eight weeks for a larger MVP. Model costs stay on your own provider account — usually $30 to $150 a month, paid directly, no markup — and you own the repository. If you are not ready to talk to anyone, the free automation audit is a three-minute questionnaire returning an automation health score, the annual cost of your manual work, and ranked quick wins. No call, no cost.
Frequently asked questions
Is Claude Code better than Cursor?
Different categories, so it depends who is driving. Claude Code is better when you describe an outcome and let an agent work across many files, run tests and return a pull request. Cursor is better when a person edits continuously. If you need model choice across vendors, Cursor is the clearer answer, because Claude Code runs Claude models only.
Who owns Cursor now?
SpaceX. Cursor's own blog announced on 14 August 2026 that "Cursor has officially been acquired by SpaceX," completing a process that began with a model-training partnership with SpaceXAI, the division formed when SpaceX absorbed xAI. Cursor's parent company is Anysphere. Practically, Cursor still offers third-party models as of September 2026 and sells the same plan tiers — but check the current model list rather than assuming it is fixed.
Can I use Codex CLI for free, and is Claude Code free?
The Codex CLI software is free and open source under Apache-2.0, but the model usage behind it is not: access comes by signing in with a ChatGPT plan, or with an OpenAI API key. Claude Code has no free tier of its own — it is included with the paid Claude plans, or billed per token against an Anthropic Console API key or cloud provider account.
What is the difference between Codex CLI and Codex cloud?
The Codex CLI runs on your machine against your local checkout. Codex cloud runs tasks in OpenAI-hosted environments, which lets you fire off several long jobs in parallel and check back later. Claude Code has the same local-versus-web split. The governance implication is identical: cloud execution means your repository is checked out on someone else's infrastructure, so read the isolation documentation before enabling it.
Is the current Codex the same as the 2021 Codex?
No — the name was reused. The original Codex was a code-completion model OpenAI released in 2021, best known for powering the first version of GitHub Copilot. Today's Codex is an agent product built on OpenAI's current GPT-5 series models, and it began rolling out in 2025. If you evaluated the original and were unimpressed, that has no bearing on the current product.
Do I need to know how to code to use these?
To produce something that runs, no. To produce something you can safely operate as a business, in practice yes — or you need someone who does. The gap is not writing code. It is reviewing it, knowing when the agent has confidently done the wrong thing, and fixing it when it breaks in a way the agent did not anticipate. Treating generated output as production-ready without review is where people get hurt.
Does Claude Code upload my code, and does it work offline?
It sends the context it needs to a model over the network, so it does not work offline — and neither does Codex or Cursor. Where that context goes depends on how you authenticate: Anthropic's API, your own cloud tenancy, or a gateway you host. Local sessions execute files and commands on your machine; cloud sessions check the repository out on the vendor's infrastructure. If no code may leave your premises at all, none of these three is the right tool.