AI model choice matters less than the coding tool used

The model vs. the tool

A recent experiment tested three popular AI coding tools—Claude Code, Codex, and Cursor—using the exact same underlying brain: Claude Sonnet 5.5 at high reasoning. The goal was to see if the specific software interface changed the quality of the code produced by the AI model. The researcher provided all three tools with the same buggy WebSocket proxy and a 30-minute window to fix it.

The results showed that the model was doing the heavy lifting for the logic. Every single tool scored a perfect 16 out of 16 on the final code tests. Even the naming conventions for functions remained consistent across all three platforms, suggesting a distinct "fingerprint" of the Sonnet 5.5 model rather than a quirk of the individual tools.

Despite the identical scores, the tools behaved differently in their execution. While the final output was nearly the same, the path taken to get there varied based on the tool's specific design. This suggests that the "agentic" behavior—how the AI plans and executes tasks—is heavily influenced by the wrapper software.

Testing the limits

The test task involved a Kindle dashboard for Home Assistant that had been ported to a specific device. The project already contained four bugs that needed fixing, and the user wanted to add an auto-reconnect feature for the proxy server. Because the Kindle is an older device, it only supports ECMAScript 5, which means the tools could not rely on modern JavaScript conveniences and had to be very precise.

To further challenge the tools, a second prompt was added to request new features, including Home Assistant status messages and an offline badge to alert the user if the server went down. The researcher also used a 16-point test routine that was kept out of the folders the agents could see. This prevented the tools from simply writing code to pass a test they could see without actually fixing the intended issue.

All three tools passed the testing suite, but they made different choices about how to handle the existing code. Codex chose to patch the original code, while both Cursor and Claude Code restructured it into smaller methods. Claude Code was the most aggressive, stripping the original code back to 34% of its previous size. This shows that while the model provides the logic, the tool determines the level of refactoring.

Setup and cost hurdles

Getting all three tools to use the same model required some technical maneuvering. Since none of the tools have a native version of Sonnet 5.5, the researcher had to use translation layers. The researcher wanted to stick to documented official routes, which pointed toward specific configurations. Cursor requires a subscription for third-party models, while Codex uses OpenRouter as an endpoint. Claude Code was restricted to Claude models to ensure the runs would actually work.

Configuration issues also emerged during the testing phase on Windows. Codex required specific TOML settings to ask for approval, and Claude Code needed a flag to use PowerShell instead of its preferred Bash environment. These hurdles highlight that the experience is not just about the AI's intelligence but also about the technical friction of the software setup. The researcher noted that the tools needed some setup time just to behave correctly on the operating system.

Pricing also varied across the different platforms. It was expected that Claude Code would be the cheapest option for using Sonnet 5.5. However, the researcher found that Cursor might have used token caching, as it cost $1.11 for roughly 3.31 million tokens, while Codex/OpenRouter cost $3.73 for 1.64 million tokens. These costs reflect how different tools manage the underlying API calls and data storage.

Workflow and experience

The researcher compared the experience of using these tools to driving a rental car. Even with the same driver and route, the fuel economy and ergonomics differ. Similarly, while Sonnet 5.5 wrote almost the same code regardless of the tool, the overall experience felt different. The tools split in how they handled speed, self-checks, and specific quirks, making it difficult to say objectively which one was superior.

Claude Code was noted for its high level of self-checking, running 42 tests to ensure accuracy by even putting the original bug back into the code to verify the fix. In contrast, Codex was much more conservative, running only 10 tests. Cursor fell in the middle with 33 tests, but it missed some features because it queued the next command by default instead of running it immediately.

Ultimately, the choice of tool depends on personal preference for specific workflows. A user might prefer Codex for minimal diffs and speed, Claude Code for an agent that aggressively checks its own work, or Cursor for a familiar IDE environment despite a potentially higher cost. The experiment proves that the software wrapper is just as important as the model it hosts.