Claude Code vs Codex vs Cursor: features, prices and limits compared (October 2026)

Anthropic's Claude Code, OpenAI's Codex and Cursor are the three coding agents most developers end up choosing between. A year ago they were easy to tell apart: a terminal tool, a cloud tool, an editor. They've since converged. All three now offer a command-line agent, IDE integration, cloud agents that open pull requests, Model Context Protocol (MCP) support and AGENTS.md instruction files. The differences that matter now are less visible. They come down to how each one meters your usage, which models you can use, how commands are sandboxed, and where your code goes.

This comparison is based on official pricing pages and documentation, checked on October 8 and 9, 2026, and on published evaluations where they exist. For why public leaderboards can't tell you which agent suits your codebase, see our coverage of published coding-agent evaluations.

At a glance

Last verified: October 9, 2026. The underlying rows, with a source URL for each, are in our data file A5-coding-agents-plans.csv.

Who each is for

Claude Code suits developers who already pay for Claude and want a terminal-first agent with the most documented ways to customize it. That means CLAUDE.md memory, skills, settings hooks, subagents and, since October 1, in-process "mods" (see our mods explainer). The trade-off is that you only get Claude models.

Codex suits ChatGPT subscribers who want a coding agent included in a plan they already have. It has OpenAI's newest models and the most restrictive documented local sandbox defaults of the three. Our Codex explainer covers how it works in detail.

Cursor suits developers who live in an IDE and want to switch between models from several vendors, including Claude and OpenAI models, in one tool. It also has the most integrations for kicking off cloud agents. For what changed in its latest release, see our Cursor news story.

Pricing

Claude Code (included in Claude plans)

Last verified: October 9, 2026, against Claude's pricing page. The page lists Max as "from $100/month"; the $100 and $200 tiers come from Anthropic's help center.

Anthropic's pricing page says Claude Code "shares the same usage limits as the rest of your plan." That means a long chat in the Claude app eats into your coding budget. Max plans have two weekly limits, one for all models and one for Sonnet only, and both reset seven days after your session starts. When you hit a limit, Pro and Max subscribers can keep working by buying usage credits at standard API rates.

In its October 7, 2026 Haiku 5.5 announcement, Anthropic said it would roll out a monthly API credit that week: $100 for Max 5x, $200 for Max 20x and up to $500 pooled for Team. Anthropic says the credits are for experimenting with "building tools, apps, and agents that call our API." The announcement doesn't mention Claude Code subscription usage.

Codex (included in ChatGPT plans)

Last verified: October 9, 2026, against OpenAI's Codex pricing page. The Go price matches OpenAI's January 16, 2026 announcement ("$8 per month" in the US).

OpenAI publishes estimated local messages per five hours for Plus and standard Business seats:

• GPT-6 Astra: 5–45
• GPT-6.1 Sol: 15–160
• GPT-6 Sol: 15–150
• GPT-6 Luna: 350–3,000

OpenAI says these are estimates, weekly limits may also apply, and cloud tasks "may use more of your allowance than local messages." Fast mode uses 2.5x your included usage.

Cursor

Cursor splits usage into two pools:

• Cursor models: Grok and Cursor's own Composer, with "significantly more included usage."
• Other models: third-party models such as Claude and GPT, billed at each model's API rate.

Once you've used up the included amount, Cursor says on-demand usage "is billed monthly at the same rates" and requests are "never downgraded in quality or speed." On Teams and Enterprise, third-party model requests carry an extra $0.25 per million tokens. Cloud agents are charged at API pricing for the selected model. Cursor doesn't publish dollar amounts for what each tier includes, which makes it harder to predict your monthly bill than with the other two. Its docs offer a rough guide instead: daily Agent users typically spend $60–$100 a month, and power users often $200 or more.

Last verified: October 9, 2026. Cursor's pricing page groups Pro, Pro+ and Ultra under "$20 / mo." and Teams under "$40 / user / mo." The per-tier prices above come from Cursor's pricing documentation. Annual prices aren't shown.

How the usage limits actually work

The meters work differently:

• Claude Code and Codex cap how much you can do in a rolling five-hour window, with a weekly cap on top. Neither Anthropic nor OpenAI publishes a fixed token number. Anthropic describes limits as multiples of other plans ("5x Pro"). OpenAI gives message ranges that vary with task size: on Plus, its estimate for GPT-6 Astra is as low as 5–45 local messages per five hours. That's why both companies sell $100–$200+ tiers aimed at heavy coders.
• Cursor behaves more like a prepaid API account. The bill depends on which model you choose and how much context each request carries. Picking an expensive model or a large context window drains your allowance faster, and you then pay on demand.

The practical upshot:

• Predictable cost: Claude Code and Codex. When the window runs out, you wait or buy credits.
• Predictable access: Cursor. You can keep going, but your bill varies.

Capabilities compared

Where they run

Claude Code

• Runs in the terminal, VS Code and JetBrains, the Claude Desktop app's Code tab, the web at claude.ai/code, the mobile app, and Slack.
• claude --cloud sends a task to an Anthropic-managed VM. claude --teleport pulls a cloud session back into your terminal.
• Cloud sessions are available on Pro, Max and Team, and on Enterprise premium or Chat + Claude Code seats. Anthropic says they share rate limits with your other usage, with "no separate compute charge."

Codex

• Runs in a desktop app, on the web, in the CLI, in an IDE extension, on iOS, and as cloud tasks.
• With an API key you get the CLI, SDK and IDE extension, but no cloud features.

Cursor

• Runs in the Cursor IDE and a CLI.
• Cloud agents (formerly Background Agents) run "in isolated VMs in the cloud." You can start them from cursor.com/agents, the iOS app, Slack, GitHub or Bitbucket comments, Linear or the API.

Models

• Claude Code runs Claude models only. On the Anthropic API, the opus, sonnet and haiku aliases point to Opus 5.5, Sonnet 5.5 and Haiku 5.5, and the default model is Opus 5.5. Fable 5.1 is opt-in (/model fable). Anthropic says Fable "can bill to usage credits instead of drawing on your plan's included limits," depending on plan. Through Amazon Bedrock, Google Cloud or Microsoft Foundry, the aliases resolve to older models in some cases.
• Codex runs OpenAI models only. On Plus, that's GPT-6.1 Sol and GPT-6 Luna. GPT-6 Astra appears in the usage table, and Pro adds Astra Ultrafast at the $500 tier.
• Cursor's model page lists Claude Opus 5.5, Sonnet 5.5, Fable 5.1 and Haiku 5.5, plus OpenAI's GPT-5.6 Sol, Terra and Luna. It also lists Gemini 3.1 Pro and 3.8 Flash, Grok 4.5–4.7, Composer 2.5 and Muse Spark 1.3. No GPT-6 model appeared in that list on October 9, 2026, and Cursor's changelog entries since August 27 don't mention one, though the pricing table is truncated behind a "Show more models" control. Check in the app before you buy for a specific model.

Permissions and sandboxing

Codex has the most conservative defaults:

• Its sandbox modes are read-only, workspace-write and danger-full-access.
• The "Auto" preset is workspace-write with on-request approvals, and local network access is off by default.
• The sandbox uses Seatbelt on macOS, bwrap plus seccomp on Linux, and a native Windows sandbox.

Claude Code relies on permission modes rather than an OS sandbox by default:

• From v2.1.283, interactive terminal and VS Code sessions start in auto mode, where a second model (a classifier) reviews actions instead of prompting you. You can switch to Manual mode, which asks before edits and commands, with Shift+Tab, and organizations can turn auto mode off.
• Its OS-level Bash sandbox is off by default. Turn it on with /sandbox. It uses Seatbelt on macOS and bubblewrap on Linux and WSL2, and it isn't available on native Windows.
• The sandbox covers shell commands only. File tools, MCP servers, hooks and mods run outside it.

Cursor runs terminal commands "in a restricted environment that blocks unauthorized file access and network activity." When commands enter that sandbox depends on your Run Mode setting.

For why these defaults matter, see the case against unattended agents.

Extensibility

All three support MCP and AGENTS.md. Our guide to writing CLAUDE.md and AGENTS.md files explains how each one loads them, including the precedence trap in Claude Code. Beyond that:

• Claude Code also has skills, settings hooks, subagents, plugins and, new this month, in-process mods.
• Codex reads AGENTS.override.md and supports approval reviewers.
• Cursor has .cursor/rules with four application modes.

GitHub integration

• Claude Code: GitHub Actions and GitLab CI integrations, an automatic Code Review product, and Auto-fix. Auto-fix watches a pull request and pushes fixes for CI failures and review comments. It requires the Claude GitHub App.
• Codex: comment @codex review on a pull request for a review that flags only P0 and P1 issues, or turn on automatic reviews. Codex reads ## Code Review Rules sections in AGENTS.md.
• Cursor: comment @cursor on a pull request or issue to start a cloud agent that pushes a branch and opens a pull request. Bugbot provides automated reviews.

Privacy and data

• Anthropic: Free, Pro and Max users choose whether their data, including Claude Code sessions, can be used for training. Retention is five years if they opt in and 30 days if not. Team, Enterprise and API usage isn't used for training under commercial terms. Zero data retention is available to qualifying Enterprise accounts.
• OpenAI: says it doesn't train on ChatGPT Business, Enterprise, Edu or API data by default. Its enterprise privacy page doesn't mention Codex by name, so check your workspace's data controls.
• Cursor: says that with Privacy Mode on, "code data is not used for training by us or our model providers." Teams get team-wide privacy mode.

What about benchmarks?

As of October 9, 2026, no independent, like-for-like evaluation of these three products as shipped had been published. The closest public sources measure something narrower:

• Terminal-Bench ranks model-and-agent combinations by the share of tasks resolved; its maintainers announced version 4.0 on August 28, 2026. Its live leaderboard is the place to check current entries, but each result reflects one model, one agent harness and its settings, so it isn't a product comparison.
• SWE-bench Pro's public leaderboard, run by Scale AI, scores model plus harness combinations, mostly with the mini-swe-agent harness. It doesn't currently list the models these products use by default.
• OpenAI said in February 2026 that SWE-bench Verified "no longer measures frontier coding capabilities," citing flawed tests and training-data contamination.

Vendor-published scores measure a model in the vendor's chosen harness, not the product you'd use. We cover what each of these evaluations measures, and its limits, in our coverage of published coding-agent evaluations.

Strengths and weaknesses

Claude Code

• Strengths: the most documented extension mechanisms (memory, hooks, skills, subagents, mods); strong cloud-to-terminal handoff; Auto-fix for pull requests; a single subscription covers chat and coding.
• Weaknesses: Claude-only models; Bash sandbox is opt-in; limits are shared with chat use; mods add an unsandboxed trust surface.

Codex

• Strengths: most restrictive local sandbox defaults (on by default, network off); included from the $20 Plus tier (with limited access even on Free and Go); OpenAI publishes approximate message ranges; @codex review built in.
• Weaknesses: OpenAI-only models; the published ranges are wide (GPT-6.1 Sol: 15–160 messages per five hours); the top Pro tier costs $500.

Cursor

• Strengths: multi-vendor model choice; an IDE built around the agent; the most documented ways to start cloud agents; usage keeps flowing past the allowance.
• Weaknesses: costs are harder to predict; included amounts aren't published; Teams costs $40 per user against $25 for the other two (monthly billing); a $0.25 per million token surcharge on third-party models for teams.

Key differences

1. Model lock-in. Claude Code and Codex each tie you to one lab's models. Cursor doesn't.
2. Metering. Claude Code and Codex use time windows. Cursor uses spend.
3. Sandbox defaults. Codex's is on, Claude Code's is opt-in, and Cursor's depends on Run Mode.
4. Customization depth. Claude Code documents the most extension mechanisms, especially since mods.

Recommendation by use case

• You already pay for ChatGPT Plus: start with Codex, since it costs you nothing extra. Upgrade only if you hit the five-hour window regularly.
• You already pay for Claude Pro: start with Claude Code. If your sessions run for hours, Max 5x ($100) is the realistic tier.
• You want one tool with every model: Cursor Pro. Watch the on-demand meter in your first month.
• Security-sensitive repository, local work: Codex's defaults, or Claude Code with /sandbox turned on and no unreviewed mods.
• Team of 5–20, pull-request-centric workflow: compare Claude Team Standard ($25 monthly) with Codex on ChatGPT Business ($25 monthly) on your own tickets. Cursor Teams costs $40 but includes Bugbot reviews and shared cloud agents.
• You want to automate review on every pull request: all three can. Run a two-week trial on real pull requests before you commit.