Three color-coded AI coding assistants connect to a central development dashboard in a modern workspace.
A new hands-on test from XDA-Developers writer Joe Rice-Jones asks a question many developers assume they already know the answer to. If you run Claude Code, OpenAI's Codex and Cursor on the same model, do you get the same work back? On the code that passed, mostly yes. Almost everything around the code differed: speed, self-testing, how much existing code was rewritten, setup on Windows, and the bill.

This is one author running one project, so treat the numbers as an anecdote, not a benchmark. It is still a useful data point for anyone choosing an AI coding agent for Windows development.

Three color-coded AI coding assistants connect to a central development dashboard in a modern workspace. The setup: one model, three harnesses​

The author gave all three tools Claude Sonnet 5.5 at high reasoning, a buggy WebSocket proxy that feeds Kindle data from Home Assistant, 30 minutes and $10 each. The project is a Kindle dashboard for Home Assistant, ported from a version for a supported Kindle Paperwhite to the author's own, older device. It had four known bugs, and the author also wanted an auto-reconnect feature.

The target device adds a constraint. The Kindle's browser only supports ECMAScript 5, so modern JavaScript conveniences are off the table. A second prompt then asked for Home Assistant status messages and an offline badge.

The author scored the results with a 16-test routine kept outside the folders the agents could see. The stated reason is that an agent can see a visible test and write code that passes it without working as intended.

How the model was matched​

Getting the same model into all three tools took some work. The author wanted documented official routes. They say Cursor needs a paid subscription to use third-party models, and Codex has a supported route through OpenRouter. They also say Claude Code is built to talk to Claude models, and other models would add an unguaranteed translation layer.

That last point is why Claude Code stayed on Claude. Cursor offered Sonnet 5.5 via API credits, and Codex supports OpenRouter, so API calls could be used on both. Cursor and Codex defaulted to high reasoning. Claude Code defaulted to medium, so the author raised it.

Anthropic's own announcement is consistent with that last detail. It says the default effort is Medium in Claude Code and the apps, while the Claude Platform defaults to High.

The model itself is real and recent. Anthropic's announcement is dated September 28, 2026. It prices Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million tokens. Anthropic cut the cache-read rate to $0.10 on October 7, 2026. The article does not give a run date, so the author's dollar figures are a snapshot, not current universal rates.

Results: everyone passed, nobody was identical​

All three tools scored 16/16 on the author's hidden tests. The differences lie elsewhere.

ToolAgent-written testsOriginal taskSecond promptTotalTokensReported costLines added / removed
Codex109:001:07about 10:071.64M$3.73+754 / −74
Cursor3311:241:37about 13:013.31M$1.11+1,218 / −164
Claude Code42 (change suite run three times)7:257:08about 14:332.55Mabout $0.96 (API-equivalent)+1,179 / −176

All figures are the author's. The costs are not a normalized comparison. They mix subscription fees, API rates and cache accounting, and the Claude Code figure is an API-equivalent estimate, not an invoice.

Where the model's fingerprints showed​

When the task was small and clear, the code converged. The author says Codex and Claude Code used the same function name for the dashboard change. The same pattern appeared in the reconnect fix. The author reads this as a trait of Sonnet 5.5 and not of the harnesses. That is a reasonable reading, but it comes from one project.

All three tools also added the same unrequested extras. None of these were in the TASK.md planning file:

  • A banner shown when the Kindle loses its connection to the proxy.
  • A fix for a malformed-JSON crash.
  • A fix for a return-versus-continue bug.
  • A check of proxy status on first connect.

Where the harness changed the outcome​

The tools diverged on open-ended judgment calls. These were how much of a file to restructure, how much defensive code to add, and how much testing to write.

  • Codex patched the proxy module and kept most of the original code. It produced the smallest diff, and it was the fastest overall.
  • Cursor restructured the proxy into smaller methods and said in its summary that it rewrote the file. The author says it kept 43% of the original.
  • Claude Code also restructured the module and kept 34%. It ran its test suite three times. It also reintroduced the original bug to confirm that its new tests would catch it. The author notes this probably didn't matter on a project this small. It could matter on larger, messier codebases.

One quirk was the language. The original repo was in German. The author says Codex and Cursor used German when rewriting the UI text, while Claude Code used English. The author declines to guess why, and notes you can simply tell an agent which language you want.

Windows setup friction​

The author says all three tools needed adjustments on Windows. These are one person's experiences, and the write-up gives no version numbers, key names or flag values.

  • Codex: it reportedly ignored the --ask-for-approval flag until the setting was added to its TOML config file. The server then crashed because it couldn't launch PowerShell without an environment value being set.
  • Claude Code: it prefers Bash, so the author set a flag to make it use PowerShell instead of trying WSL.
  • Cursor: it opened the run in ask-for-approval mode even though the author had set it to auto.

For Windows teams, the practical point is to test approval and shell behavior before trusting an agent with unattended work. Check which shell it launches and which permission mode it is really in. Don't assume your settings were honored.

OpenAI's engineering write-up on the Codex agent loop gives some background for the config-file behavior. It is not quoted here, but it describes the Codex CLI as reading its settings from a config.toml file. It also says the Responses API endpoint is configurable, so any endpoint that implements that API can be used. That fits the author's choice of an OpenRouter route. It does not confirm the exact configuration used in this test.

Steering and cost: the caveats​

The steering test didn't happen​

The author wanted to see how each tool handled a second prompt while still working. Claude Code and Codex finished the first task in under ten minutes, so only Cursor was mid-task. Cursor queued the new command by default and did not steer the running agent. The author kept defaults and did not change this. So the steering comparison is asymmetric and shouldn't count as a result.

Why Cursor's bill was low​

Cursor's $1.11 for about 3.31 million tokens, against $3.73 for 1.64 million on Codex through OpenRouter, looks odd. The author's explanation is that most of Cursor's token reads were cached. The author also says Codex appeared to pay full price, with no cache reads. That is an inference. It is not verified billing data, so don't generalize it. The author adds that third-party models in Cursor needed a paid plan, which they put at $30, plus API billing. Plan terms change, so check current pricing.

How this fits with other testing​

Other same-model tests found different gaps, which is a reminder that results depend on the task. A write-up on DEV Community describes a test by a Dutch site, notanaiapp.com. It reports that both tools passed all 186 hidden tests every time. It also reports Claude Code sending roughly 269,000 tokens per job against about 92,000 for Cursor. That is the same kind of finding as XDA's. Correctness matched, and cost and behavior did not.

Endor Labs, on the other hand, tested Sonnet 5 (not 5.5) with Cursor. It reported that Cursor scored 63.1% FuncPass against Claude Code's 83.2%, with the gap traced to a Cursor-specific throughput problem for that model. Different model, different task and different metric, but it shows that the harness can matter a lot in some conditions.

What to take from it​

  • Matching the model doesn't make the tools interchangeable. The gaps were in diff size, self-verification, interaction defaults, setup and cost.
  • Small, well-defined jobs converge. Expect similar code from the same model whichever tool wraps it.
  • Ambiguous jobs diverge. Refactor scope, defensive code and test volume are where your choice of tool shows up.
  • Keep your acceptance tests out of the agent's reach. The author's reason is sound. An agent that can see the tests may write to them.
  • Verify the setup on Windows. Confirm approval mode, shell choice (PowerShell versus Bash or WSL) and config-file settings before a long run.
  • Don't generalize the costs. The prices depend on plan, caching, API route and date.

The author's own preferences are personal. Codex suits minimal diffs and speed. Claude Code suits an agent that checks its own work. Cursor suits an IDE workflow, at a higher overall cost. The author feels most at home in Claude Code but does most agentic coding in a different tool, Crush, because Claude Code gives summaries instead of detailed planning. Nothing here shows that one tool is generally better. One project and one 16-test suite can't settle that.

The result that matters for IT teams is that all three produced working code, so the choice comes down to workflow, controls and cost. If you are standardizing on one, run the same kind of small bake-off on your own codebase. That means your own tests, your own shell, and your own rules for approvals.


Correction (October 22, 2026): The original text said Anthropic’s September 28, 2026 announcement priced Sonnet 5.5 cache reads at $0.10 per million tokens. The announcement price was $0.20; Anthropic cut the cache-read rate to $0.10 on October 7, 2026.

 

References

  1. I ran Claude Code, Codex, and Cursor on the same AI model, and the tool mattered more than I expected XDA 2026-10-22T21:31:20+00:00
  2. Unrolling the Codex agent loop openai.com
  3. Introducing Claude Sonnet 5.5 \ Anthropic anthropic.com