Cursor vs Claude Code vs GitHub Copilot vs Codex: A Full Comparison — How to Choose an AI Coding Tool

Cursor, Anthropic's Claude Code, GitHub's GitHub Copilot, and OpenAI's Codex are all coding agents: they read your repository, write code, run commands, and carry the work through to a finished change on their own. It used to be possible to tell them apart by type: "Cursor is an editor, Copilot is a plugin, Claude Code lives in the terminal, Codex runs in the cloud." But when we reread all four vendors' official documentation on September 22, 2026, all four run in almost every place: the terminal, the IDE, a desktop app, the cloud, and your phone.

So "where does it run?" is no longer the question to ask when choosing. What matters now comes down to three things: (1) which vendors' models you can use, (2) how pricing and usage are counted, and (3) where each tool's home base is (its own editor, GitHub, a Claude subscription, or a ChatGPT subscription).

Updated on October 16, 2026: Cursor moved to The X Subscription, a single plan bundled with Grok and X, from October 12, so we updated Cursor's pricing, how usage is counted, and its organization plans.

Fully rewritten on September 22, 2026. Statements in the previous version such as "the four are different types," "Copilot is the cheapest," and "OpenAI claims Codex is about 4x more token-efficient than Claude Code" have been corrected or removed, because they don't match current official information or couldn't be verified in primary sources. What we changed and how is listed at the end of the article.
Updated on September 26, 2026. Following the arrival of Claude Opus 5.5 (September 22) and GPT-6 Sol and Luna (September 23), we revised the model list and Claude Code's default model in section 4, and added that OpenAI plans to stop providing its models to Cursor on November 12, 2026.

1. The answer: the differences are in models, pricing, and home base, not where they run

Here's the whole picture on one page first. Sections 2 onward check each condition and source one at a time.

In most cases the choice settles itself if you look, in this order, at (1) the subscription you already pay for, (2) the models you want to use (whether you'll commit to one vendor or move between several), and (3) where the center of your work is (the editor, GitHub, or the terminal). Answers for specific scenarios are in section 7.

2. "Editor / plugin / CLI / cloud" no longer works as a way to sort them

Over 2025 and 2026, each of the four vendors filled in the entry points it was missing. Here are only the ones whose dates can be confirmed from official sources.

Sources: GitHub Blog (April 4, 2025), GitHub Changelog (July 16, 2025), same (September 25, 2025), same (February 25, 2026), same (June 17, 2026), same (July 7, 2026) / Cursor, "Agents Window" / Codex, "What's new"

As a result, lined up by entry point, the picture looks like this.

Sources: Cursor, "CLI", Cloud Agents, Cursor for iOS, JetBrains / Claude Code, "Platforms and integrations", Use Claude Code in the cloud / GitHub Docs, "Plans for GitHub Copilot", About Copilot cloud agent, Usage-based billing for individuals (list of supported IDEs) / Codex, "IDE extension", Pricing (all checked on September 22, 2026)

The four are nested inside one another

There's one more thing that breaks the old "four types" framing: the four aren't mutually exclusive products. They run inside one another.

• Claude Code and Codex run inside Cursor: both IDE extensions list Cursor as a supported editor
• You can call Claude and Codex from within Copilot: on GitHub, you can assign an issue to a third-party agent (Anthropic Claude or OpenAI Codex) (public preview)
• Cursor and Copilot let you choose Claude and GPT models: the same Claude Opus is available from Cursor, Copilot, or Claude Code (in Copilot, on Pro+, Max, Business, and Enterprise; the $10 Pro plan can't select Opus models)

Sources: GitHub Docs, "About third-party coding agents", Claude Code, "VS Code"

In other words, the question "which one should I pick?" actually splits into three: which screen you'll work in, which model you'll use, and whose pricing you'll pay under.

3. Home base: where each one is most natural to start

Even though they all run in nearly the same places, each vendor has polished a different place the most. That's what makes the difference in effort when you start using them.

Sources: Cursor, "Migrate from VS Code", Models & Pricing / GitHub Docs, "About Copilot cloud agent" / Claude Code, "Platforms and integrations" / Codex, "Pricing", Open Source

Cursor's own features and pricing are covered in depth in our Cursor explainer, and Copilot on its own in our GitHub Copilot explainer. If you want a detailed comparison of just Claude Code and Codex, Claude Code vs Codex goes as far as sandboxing and computer-use conditions.

4. Available models: two that span vendors, two that use only their own

This is the biggest dividing line among the four. Cursor and Copilot can switch between models from multiple vendors, Claude Code runs only on Claude models, and Codex runs only on OpenAI models.

Sources: Cursor, "Models & Pricing", Cursor Router / Claude Code, "Model configuration" / GitHub Docs, "Plans for GitHub Copilot", Usage-based billing for individuals, Models and pricing / Codex, "Models", Pricing (all checked on September 26, 2026)

The deciding factor is whether the model you want is settled on one vendor. If it is, that vendor's own tool (Claude Code or Codex) is the straightforward choice, since it runs directly on your subscription's allowance. If you want to move between vendors, or try every new model as it comes out, Cursor or Copilot gets you there with a single subscription. For details on each model, see our articles on Claude Opus 5.5, GPT-6 Astra, and GPT-6 Sol and Luna; for every vendor's current lineup, see knowledge cutoff dates for major models.

OpenAI's models are scheduled to stop working in Cursor

Cursor was acquired by SpaceX on August 14, 2026. In response, on August 28 OpenAI notified SpaceX that it intends to end its contract to supply models to Cursor, proposing November 12, 2026 as the cutoff date. OpenAI also said it won't provide new models to Cursor in the meantime. November 12 is only OpenAI's proposal: its help article says it will share the official termination date once both companies confirm it, and that Cursor may choose to end access sooner. That's why the GPT-6 family is missing from Cursor's model list as of September 26, 2026. If you mainly use GPT models in Cursor, once access ends you'll need to switch to Claude, Gemini, or Cursor's own models, or use GPT through Codex or Copilot instead. To keep using them inside Cursor, OpenAI's help also points to bringing your own OpenAI API key, which works in Cursor's local Chat and Agent and is billed at API prices (sources: OpenAI, "Our decision on Cursor following its acquisition by SpaceX", Cursor, "Cursor is now a part of SpaceX", OpenAI Help, "Using OpenAI models in Cursor").

Same model, different allowance

When you choose a Claude model in Cursor or Copilot, it's Cursor's or Copilot's allowance that goes down. Cursor doesn't count other vendors' models toward its weekly allowance. On Premium and higher plans, it deducts them at each vendor's API list price from prepaid credits (Plus and Free can't use other vendors' models). Copilot likewise deducts AI credits at each model's per-token rate. For example, Claude Opus 5.5 is charged at $4 per million input tokens and $20 per million output tokens in both Cursor and Copilot (Cursor as of October 16, 2026; Copilot as of September 26, 2026). It doesn't use your Claude Pro or Max allowance, or your ChatGPT Plus or Pro allowance.

Why context length is out of the comparison table

The previous version's table had rows like "200K context window (up to 1M with Max Mode)" and "about 400K." But the usable length is set by the model and plan, not the tool, and it changes every time the model does. Cursor's Max Mode is now available only on the legacy request-based pricing plans. For those reasons, we took it out of the tool comparison.

5. Pricing: free tiers, entry plans, top plans, and going over

Here are the monthly prices for individuals, arranged by price tier.

Cursor has been on The X Subscription's five tiers since October 12, 2026. The older Hobby, Pro ($20), Pro+ ($60), and Ultra plans continue only for people already subscribed. Annual billing is Plus $80, Premium $300, Super $1,000, and Ultra $2,000 (about 17% off the monthly price). Sources: Cursor pricing page, Cursor "Models & Pricing", SpaceXAI "Introducing The X Subscription" (Cursor checked October 16, 2026) / Claude pricing page / GitHub Docs, "Plans for GitHub Copilot" / Codex, "Pricing", OpenAI Help, "About ChatGPT Pro tiers" (Cursor checked October 16, 2026; others on September 22, 2026)

"Cheapest" flips depending on how you compare

Compared on monthly price alone, the cheapest paid options are Cursor Plus and Codex Go (both $8), followed by Copilot Pro ($10), and Claude Pro is $20. However, Cursor Plus only gives access to Cursor and Grok models; to use Claude or GPT in Cursor you need Premium ($30). That's why the previous version's "Copilot is the cheapest" was wrong.

However, how much work you get for what you pay can't be compared from this table, because the four vendors don't measure their allowances in the same units.

• Copilot: publishes its monthly allowance as a dollar amount. Pro is 1,500 credits ($15 worth), of which 500 are a bonus portion that varies with model pricing and efficiency. Pro+ is 7,000 and Max is 20,000. Credits are drawn down at each model's API list price
• Cursor: shows the weekly allowance for its own models (Grok and Composer) as a multiple of Plus (Premium 4x, Super 15x, Ultra 50x), not as a dollar amount. Other vendors' models are outside it and are paid from prepaid credits at API list price
• Claude and Codex: don't express the allowance in dollars. OpenAI publishes guideline figures for messages per 5 hours by model and plan, but notes they aren't fixed limits

A more useful reference is who each vendor itself recommends each plan for. Cursor recommends Premium for people who are starting to use agents, Super for people who use them every day, and Ultra for people who run multiple agents or automations. OpenAI describes Plus as suited to a few focused coding sessions a week; Anthropic positions Max for people who work with Claude all day; and GitHub positions Copilot Max for sustained heavy use. If you plan to hand work to an agent every day, the realistic approach is to assume a $10–$20 entry plan won't be enough, and to weigh a higher plan while watching your usage screen for a month or two.

Four ways of counting limits

Sources: Cursor, "Usage and limits" / Anthropic Help, usage credits / GitHub Docs, "Usage-based billing for individuals", GitHub Changelog (June 1, 2026) / Codex, "Pricing"

You really feel the difference when you hit a limit. With Cursor and Copilot, which count by the month, running out mid-month means either paying to continue or waiting until next month. With Claude Code and Codex, which have two tiers of 5-hour and weekly limits, a 5-hour limit comes back within a few hours, but hitting the weekly limit stops you until the next reset. Either way, you can keep going by paying for the overage. What Claude Code shows when you hit a limit, and what to do about it, is covered in our article on "You've hit your session limit".

On June 1, 2026, Copilot switched from counting use of premium models by monthly request count (premium requests) to AI credits, which convert token consumption at each model's published API rates. The base monthly prices (Pro at $10 and so on) didn't change. Annual Pro and Pro+ subscriptions stay on the old system until the subscription ends (GitHub Blog).

Rolling it out to a team or company

Sources: Cursor, "Models & Pricing" / Claude pricing page / GitHub Docs, GitHub Changelog (August 28, 2026) / Codex, "Pricing"

For Copilot Business and Enterprise, GitHub has announced that for organizations paying by card or PayPal, it is reopening new sign-ups in stages from September 1, 2026, and that from October 1 seat fees will be prepaid, including for existing organizations. When rolling a tool out at work, you'll avoid arguments later if you settle not just the price but also how you'll pay and who approves usage.

One more thing: for Copilot Business and Enterprise, GitHub added a Default policy for new features policy on September 24, 2026, which takes effect on October 22, 2026. Eligible generally available features an administrator has left unconfigured will then follow the chosen default (Enabled, Disabled, Let organizations decide). Features you have explicitly enabled or disabled keep your choice, and preview features stay opt-in. New models are also enabled automatically unless an administrator turns off the global default or disables them individually. You can pick the default before it takes effect, so organizations that want to hold certain features back can decide now (GitHub Changelog).

6. GitHub PRs and reviews: where the biggest differences remain

Even now that they all run in the same places, there's still a gap in how much of the GitHub issue-and-PR cycle you can run without leaving GitHub.

Sources: Cursor, "Cloud Agents", Bugbot, Cursor Changelog (September 23, 2026) / Claude Code, "cloud", Code Review, GitHub Actions / GitHub Docs, "Plans", GitHub Changelog (June 1, 2026) / Codex cloud, Pricing

If your team's work runs on GitHub's issue → PR → review cycle, Copilot takes the fewest steps. Pick Copilot in the assignee field and the work starts; the review happens on the same screen. And as noted earlier, you can assign Claude or Codex agents from that same field (public preview). Conversely, if your repositories are a mix of GitLab, Bitbucket, and Azure DevOps, Cursor, whose Cloud Agents and Bugbot support all of GitHub, GitLab, Bitbucket, and Azure DevOps, is easier to work with.

7. Which one to pick: deciding by eight scenarios

Here's how the conditions so far apply to common scenarios. If more than one fits, give priority to the one higher up.

Sources: Cursor, "Rules (AGENTS.md)" / GitHub Docs, "Adding repository custom instructions" / Codex, "AGENTS.md" / Claude Code, "memory"

If you've written an instruction file and it isn't being followed, how to check why is covered in why AI ignores your rules and how to fix it.

8. Using them together, and how to compare them on your own repo

Because the four run inside one another, using them together isn't hard. Here are common combinations and the reasoning behind each.

Whatever the combination, follow one rule: never let two agents write to the same working folder at the same time. They can overwrite each other's changes. If you run them in parallel, give each its own folder with worktrees. For a real workflow that goes all the way to deployment, see deploying to Vercel with Claude Code and Cursor.

A 30-minute way to compare

More reliable than benchmarks is having the tools solve the same tasks in your own repository. Try the following with the two or three of the four you're most curious about.

If no difference shows up, any of the tools will do. In that case, the fastest decision is to go by pricing and home base (sections 3 and 5). If you want to learn how to use the four for different jobs hands-on, there's also Chapter 3 of our Practical AI Coding course, "Comparing Cursor, Copilot, and Codex — and Choosing Between Them".

9. Why there's no benchmark comparison table

The previous version included phrasing like "Claude Code ranks near the top on SWE-bench Verified-type tests" and "the strongest for complex work." Like our Claude Code vs Codex comparison and our model release articles, this site's policy is to only line up figures measured by the same publisher with the same yardstick, so we removed them. In a comparison of these four tools, two points apply in particular.

For the same reason, we also removed the previous version's sentence "OpenAI claims Codex is about 4x more token-efficient than Claude Code." We couldn't confirm that claim in OpenAI's official announcements or documentation.

Summary

As of September 2026, Cursor, Claude Code, GitHub Copilot, and Codex all run in the terminal, the IDE, a desktop app, the cloud, and on your phone, and they run inside one another too. The days of choosing by type (editor / plugin / CLI / cloud) are over.

Today there are three axes for choosing. If the model you want comes from one vendor, go with that vendor's Claude Code or Codex. If you'll move between vendors, go with Cursor or Copilot. And on top of that, if your work centers on GitHub, Copilot; if you want to hand over the whole editor, Cursor. Pricing can't be compared on monthly price alone, so use a free tier or entry plan for a week or two, measure how "heavy" your use is on the usage screen, and then move up.

Related articles: Claude Code vs Codex, What is Cursor, What is GitHub Copilot, Which is better for translation, plus recommended models, Claude vs ChatGPT pricing comparison.

FAQ

Q1. So which one should I start with?

If you already pay for Claude or ChatGPT, start with the Claude Code or Codex that comes with it. If you don't pay for anything, pick from Codex, Copilot, and Cursor, which have free plans, choosing the one closest to the center of your work (Copilot for GitHub, Cursor for the editor, Codex if you use ChatGPT a lot). The order of decisions is laid out in section 7.

Q2. Isn't Copilot the cheapest?

By monthly price, the cheapest paid option is Codex, included in ChatGPT Go ($8). Copilot Pro ($10) comes next. However, the four vendors count their allowances in different units, so how much work you get for the same money can't be compared from a table. Copilot Pro includes $15 worth of AI credits at API list price, which run out faster the more you use heavy models on long tasks. Details are in section 5.

Q3. If I can use Claude models in Cursor, do I still need Claude Code?

Even with the same model, the allowance that gets used and the way the tool is built differ. Using Claude in Cursor requires Premium or higher, and API list price is paid from prepaid credits, not your Claude Pro or Max allowance. If you already pay for a Claude subscription, using Claude Code costs nothing extra. The surest way to find out which suits your work is to select the same model and compare them using the steps in section 8.

Q4. I use GPT-5.5 in Codex. What happens?

GPT-5.5 was retired from every ChatGPT, ChatGPT Work, and Codex plan on October 14, 2026. The OpenAI API and Codex signed in with an API key are not affected. OpenAI advises replacing any place where your configuration specifies gpt-5.5 with GPT-6.1 Sol (gpt-6.1-sol) on Plus, Pro, Business, Enterprise, and Edu, and with GPT-6 Luna (gpt-6-luna) in the desktop app on Free and Go (Codex, "Models", checked on October 16, 2026). Note that as of September 26, 2026, GPT-5.5 is still on the model lists in Copilot and Cursor.

Corrections to the previous version

We corrected the following statements from the previous version of this article that don't match current official information or couldn't be verified in primary sources.

Some statements were kept. Copilot Pro at $10 and Pro+ at $39 and the June 1, 2026 move to usage-based billing (AI credits); Codex's Go at $8, Plus at $20, and Pro at $100 or $200; and Claude's Pro at $20 and Max at $100 or $200 are all still correct, so we left them in.

Sources

• Cursor: Pricing page / Models & Pricing / Pricing and plans / Usage and limits / Cursor Router / CLI / Cloud Agents / Cursor for iOS / Agents Window / JetBrains / Bugbot / Rules / Migrate from VS Code / Cursor is now a part of SpaceX
• Claude Code / Anthropic: Platforms and integrations / Use Claude Code in the cloud / Model configuration / Memory / Code Review / GitHub Actions / Claude pricing page / Using Claude Code with Pro and Max / Usage credits
• GitHub: Plans for GitHub Copilot / Usage-based billing for individuals / Models and pricing / About Copilot cloud agent / About third-party coding agents / Adding repository custom instructions / Changelog: billing and plan updates (June 1, 2026) / Copilot CLI general availability (February 25, 2026) / GitHub Copilot app general availability (June 17, 2026)
• Codex / OpenAI: Our decision on Cursor following its acquisition by SpaceX / Pricing / Models / IDE extension / Codex cloud / AGENTS.md / Open Source / What's new / About ChatGPT Pro tiers
• Microsoft: FY2025 Q3 earnings call (April 2025) / FY2025 Q4 earnings call (July 30, 2025) / FY2026 Q4 earnings call (July 29, 2026)

All of the above were checked against the original text of the official pages retrieved on September 22, 2026, and the model lists, defaults, and cloud availability were re-retrieved on September 26, 2026. Pricing, plans, and availability can change, so check each vendor's page before subscribing.