Claude Code vs Cursor vs Codex: tests, benchmarks, prices

There is no single winner. In the hands-on reports we found, Claude Code gets credit for the most careful output, Cursor for the smoothest daily experience, and Codex for speed and value. On the official leaderboards, the scores belong to models, not tools, and the harness around the model changes the result. Below you get the first-hand reports, the benchmark numbers with their limits, the real plan prices, and a one-week test so you can decide on your own work instead of ours.

What are these three tools, in one minute?

They are three different ways to use a coding model, not three versions of the same product.

Sources: Claude pricing page, Cursor pricing page and ChatGPT and Codex plans. Vendors change these products often, so confirm on their pages.

What do developers who use them every day say?

These are published first-hand accounts. Each is one person's experience, on their own projects, so read them as testimonies and not as measurements.

A 5,000-hour reviewer scored Cursor 55, Codex 54, Claude Code 39

Yaroslav Dobroskok, a full-stack developer, says he has logged over 5,000 hours with AI coding tools since 2022. In a September 2026 review he scored the tools out of 60 across six criteria.

His reasons: he praises Cursor for showing the agent's work, code and diffs in one view, and says its cost rises quickly when you use outside frontier models. He rates Codex highest for value and usage per dollar. He credits Claude Code with strong models and a mature terminal agent, and marks it down for the split between terminal and editor, a buggy desktop app and tighter usage limits. He reports that one OpenAI model deleted a local database after he told it three times not to. His own setup mixes tools: Codex for personal projects and Claude Code at work, because his company chose it.

Source: NeuryCode: Cursor vs Codex vs Claude Code, 5,000-Hour Review, September 10, 2026. Scores are the author's own and subjective.

A same-prompt build test favoured Claude Code on polish and Codex on speed

A XDA Developers writer gave all three tools the same prompt to build a content-management web app, with no hints or corrections. Codex finished about three times faster than Claude Code and roughly twice as fast as Cursor. The writer scored Codex 8 of 10, Cursor 8.5 and Claude Code 9.5. Claude Code was the slowest but made sensible choices nobody asked for, such as meaningful icons, word-count templates and grouped article lists. Cursor had the most distinctive visual design. The writer's verdict was that Claude Code was closest to working like a senior developer.

Source: XDA Developers: I gave Claude Code, Codex, and Cursor the same complex web app to build. One build, one prompt and one judge, so it shows a pattern and proves nothing.

A daily user described three roles: create, continue, investigate

Sowmya Raghunathan, writing in October 2025, uses all three each day. Her summary: Codex for creating new services and large refactors, Cursor for continuity (small fixes and edits in a repo it already indexes), and Claude Code for deep debugging, design alternatives and reverse-engineering logic. She notes Claude Code needs careful setup and context, and is slower for quick iteration. She says Codex can run ahead of you, which makes it less suited to fine-grained debugging.

Source: LinkedIn: Codex vs Claude Code vs Cursor vs Windsurf, my real-world take. Written when the tools and models were older, so details may have changed.

A common pattern: Cursor to type, Claude Code to hand off

Eric Roby describes running both: Cursor for active coding and quick completions, Claude Code in the terminal for background work such as a refactor across many files, test generation or a legacy module. The two do not collide because they run in separate places.

Source: Eric Roby on Substack: Cursor vs Claude Code, why top developers are using both, January 13, 2026. More of a general workflow than a personal diary.

What a survey of engineers found

The Pragmatic Engineer newsletter surveyed 906 readers between January 27 and February 17, 2026. Claude Code ranked first for overall usage, and was the most loved tool at 46%, against 19% for Cursor and 9% for GitHub Copilot. Codex, which did not exist in the previous survey, had about 60% of Cursor's usage. People could pick several tools, and the sample is the newsletter's own readers, so it is not a random slice of all developers.

Source: The Pragmatic Engineer: AI tooling survey, March 3, 2026.

What the accounts agree on

• No one-tool winner: all four reviewers either use several tools or say they would for different jobs.
• Claude Code earns trust on quality and judgment, and loses points on speed, usage limits and daily comfort.
• Cursor earns trust on the editing experience, and loses points when external models burn through the plan's allowance.
• Codex earns trust on speed and value, with the caveat that it can race ahead of you and needs review.
• Any agent can destroy data. One reviewer above reports a model deleting a local database after being told not to. Keep backups and keep agents away from live databases, as our production checklist explains.

What do the official benchmark numbers say?

Benchmarks are useful for one thing: seeing that models differ. They are poor at telling you which tool will suit your project. Three facts first.

1. Benchmarks score models inside a test harness. A tool such as Claude Code, Cursor or Codex is the harness you use in daily life. Leaderboards often run every model in the same plain harness, so the result tells you about the model's skill, not about the tool's features.
2. The popular tests keep changing. SWE-bench Verified is now described as saturated by Vals AI, which no longer runs it on new models, and harder tests such as SWE-bench Pro have taken its place.
3. Vendor-reported scores and independent scores differ because harnesses differ. Vals AI says its numbers may not match figures published by the model makers.

SWE-bench Verified (September 1, 2026, Vals AI, same harness for all)

The top models sit within two points of each other. Vals AI states that the benchmark has saturated and that it no longer runs it on new models. Cost per test varied: Claude Opus 5 at $1.29 and DeepSeek V4 Pro at $0.10.

Source: Vals AI: SWE-bench Verified leaderboard, updated September 1, 2026.

SWE-bench Pro, the harder public test (Scale AI)

On the harder test the scores drop from the 90s to the 40s, 50s and 60s, and the order changes. The page does not show a date for this table, and the newest Claude and OpenAI models we have named elsewhere are not on it. Read it as a snapshot of how much harder this test is, and check Scale's page for the current list.

Source: Scale AI: SWE-bench Pro public leaderboard. The page lists no date and not every row names its harness.

Terminal-Bench 2.0, for command-line agent work

Here OpenAI models lead among the entries listed, and Cursor's own Composer model lands level with Claude Opus 4.7. This is an aggregator's table and the harness is not specified, so treat it as a rough guide. The official site, tbench.ai, showed an empty table when we checked.

Source: BenchLM: Terminal-Bench 2.0, updated October 20, 2026. A third-party mirror, not the benchmark owner.

How should you read the benchmark tables?

• A model that leads one test often trails on another, as the first two tables show. No model wins all of them.
• Scores of 95% and above on a saturated test cannot separate models that are within a point or two of each other.
• The newest models are not always on the official boards. Wait for the owner's page before quoting a new number.
• Your project is not a benchmark task. Hidden requirements, a messy codebase and your own review habits decide the outcome more than a three-point gap.

Do these tools actually make developers faster?

The best controlled evidence comes from METR, a research group. In a 2025 randomized trial, 16 experienced open-source developers worked on 246 real tasks in their own repositories, mostly with Cursor and Claude models of that time. Tasks where AI was allowed took 19% longer, while the developers believed AI had sped them up by about 20%.

METR now labels that result outdated. In February 2026 it ran a larger follow-up with 57 developers and more than 800 tasks. The estimates showed speedups of 18% for returning developers and 4% for new ones, but both ranges included zero, and METR called the data an unreliable signal because many developers refused to work without AI, so the study could not recruit the people most helped by it. METR's own view is that developers are likely more sped up than in early 2025, but its evidence for the size of the gain is weak.

The lesson is practical. Feeling fast and being fast are different things. Measure your own results, as the test below does.

Sources: METR: early-2025 AI and experienced open-source developers study and METR: update on the developer productivity experiment.

What do they cost?

Prices below were read from each vendor's pricing page on October 21, 2026. Usage limits are where plans really differ, and they change often.

Sources: Claude pricing, Cursor pricing and ChatGPT and Codex plans.

Our 5,000-hour reviewer estimates that near-unlimited daily use costs at least $100 a month on any of them, and suggests judging cost per finished task, not price per token. That is a sensible rule: a cheaper tool that needs three retries is not cheaper.

Which one should you choose?

Use the evidence above as a starting point, then test.

• You are new and want to see every change: start with Cursor. You see files, diffs and chat in one window.
• You like handing off a big job and reviewing the result: try Claude Code or Codex. Reviewers lean to Claude Code for careful, judgment-heavy work and to Codex for fast delivery and parallel tasks.
• You hit a hard bug or need a design second opinion: the accounts above point to Claude Code.
• You already pay for ChatGPT: Codex costs you nothing extra to try on Plus.
• You work for clients: pick one main tool, learn its limits well, and keep a second for what the first does badly. Our beginner guide to Claude Code covers the first steps.

How can you test them fairly in one week?

1. 01

   Pick two real tasks

   One small (a form field or a bug) and one bigger (a feature across several files). Use work you actually need done.

2. 02

   Use the same project copy and the same prompt

   Start each tool from the same Git branch and paste the identical instructions. Do not coach one tool more than another.

3. 03

   Record four numbers per run

   Minutes until done, number of corrections you had to make, whether the tests passed, and whether you understood the change.

4. 04

   Review the diff before accepting

   Check for secrets, deleted files and changes you did not ask for. Use Git so any run can be undone.

5. 05

   Repeat on the second task, then score

   Add up cost per finished task, not the monthly price. The tool with the lowest total effort wins for you.

Whatever you pick, review every change before you accept it, and run our production checklist before an AI-built app goes live. If you want a team to build with these tools for you, see our pricing or book a call. To learn the full workflow yourself, see the Autopilot Studio course.